Model releases and updates
View model releases and updates by region and service deployment scope. See the model deprecation policy for deprecation rules and lists.
China (Beijing)
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
qwen4-max-preview
Qwen4-Max-Preview is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
Currently in preview. Available through Token Plan only. See Qwen4-Max-Preview for details and Token Plan for access instructions.
kimi-k3.1-pro
Kimi-K3.1-Pro is a 3T-class open-weight model in Moonshot AI's K3.1 series, developed through large-scale post-training of Kimi K3. It supports a one-million-token context window, text and image inputs, text output, and native visual understanding. It is optimized for coding, 3D creation, financial research, and long-running agent tasks. It scores 65.7% on Terminal-Bench 4.0 and 64.4% on OfficeQA Pro v2, while the claim-level hallucination rate decreases from 23.4% for K3 to 8.3%.
antdigital/Ling-3.1-flash
Ling-3.1-flash is a hybrid thinking model directly provided by AntDigital. It has about 560B total parameters and activates about 25B parameters per token. Its context window of up to 1M can hold longer documents, code, and task histories. The model is continuously optimized for general-purpose agents, search, everyday office work, and software development, with ongoing improvements in healthcare, finance, and materials science research. It supports only text input and runs in thinking mode by default.
qwen-audio-3.1-realtime-flash
A real-time duplex speech model with text and audio input and output, function calling, web search, and voice cloning. The context window is 262,144 tokens.
Tripo/Tripo-ImageToSplat
Tripo generates 3D Gaussian Splat models from a single image in the .splat format. Gaussian Splatting is a new 3D representation method designed for high-quality real-time rendering and photorealistic scene reconstruction. The output .splat files can be loaded with mainstream Gaussian Splat viewers and renderers.
Tripo/Tripo-Texture
Tripo AI supports text prompts and image uploads to automatically generate texture maps for untextured 3D models using AI. It supports PBR and 8K textures, delivering more refined, high-fidelity material details and significantly enhancing the appearance of textures, metallic surfaces, and other fine details.
Tripo/Tripo-Segmentation
Tripo AI Mesh Segmentation automatically identifies and separates model components using AI, with no need for manual part separation or topology expertise. It offers three levels of segmentation accuracy, providing flexible options for 3D printing, game assets, animation, and other workflows.
Tripo/Tripo-SmartSegmentation
Tripo Smart Segmentation supports image or 3D model input and provides end-to-end model segmentation. It is designed for scenarios requiring more precise, semantically controlled segmentation, such as separating specific parts, controlling the segmentation range, or adjusting segmentation precision. It is recommended for workflows that require greater control and flexibility.
Tripo/Tripo-MeshCompletion
Tripo Mesh Completion uses AI to automatically complete separated model components, ensuring their integrity after segmentation and improving workflow efficiency for 3D printing, game assets, animation, and other applications.
Tripo/Tripo-Retopology
Tripo Retopology quickly reduces the polygon count of high-poly models to meet different LOD requirements while preserving the original geometry and details. It generates well-structured topology and supports quad meshes, making models easier to edit and use for animation.
Tripo/Tripo-ConvertFormat
Tripo AI Format Conversion supports fast conversion between multiple mainstream 3D model formats, automatically parsing model data and adapting it to the target format while preserving the model’s structure and material information. It meets the requirements of different 3D software, engines, and production workflows.
Tripo/Tripo-RigCheck
Tripo Auto Rig Check intelligently detects whether a model meets the requirements for automatic rigging and automatically determines the appropriate skeleton type based on the model’s structure and characteristics, providing accurate pre-checks for subsequent rigging and animation workflows.
Tripo/Tripo-AutoRig
Tripo Auto Rigging is powered by Tripo's proprietary UniRig algorithm, which intelligently analyzes model structures and automatically generates, positions, and binds skeletons to the model. It accurately identifies key motion dimensions, such as the spine’s bending path, enabling more precise bone placement and skin weight distribution, and quickly generates character models ready for animation production.
Tripo/Tripo-AnimationRetarget
Tripo Animation Retargeting provides 100 pre-built animations, allowing standard motions to be quickly retargeted to rigged models. It automatically adapts to differences in skeleton structures and character proportions, eliminating the need for manual animation data adjustments and enabling the rapid generation of ready-to-use character animations.
vanchin/deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is served directly by Vanchin. It supports native multimodal visual understanding and uses an architecture optimized for inference speed, throughput, and scalability.
kimi/kimi-k3.1-pro-highspeed
Kimi K3.1 Pro Highspeed is a high-speed version of Kimi K3.1 Pro served directly by Moonshot AI. It retains the same model capabilities and provides approximately two to three times the output speed of Kimi K3.1 Pro.
kimi/kimi-k3.1-pro
Kimi K3.1 Pro is a K3.1-series model served directly by Moonshot AI. It has approximately three trillion parameters and supports a one-million-token context window. It is optimized for coding, 3D creation, financial research, and long-running agent tasks.
MiniMax/MiniMax-M3.1-Flash
MiniMax-M3.1-Flash is a high-performance multimodal model for complex agentic tasks. It improves quality in software development and cybersecurity while balancing a model generation speed of about 170 tokens/s with lower usage costs. The model supports text, image, and video input, a 1-million-token context window, and five thinking-effort levels. Thinking is always enabled.
qwen-audio-3.1-tts-plus
Qwen-Audio-3.1-TTS-Plus is a high-quality, professional text-to-speech model for premium voices and content creation. It supports multiple languages and dialects, streaming speech synthesis, free-style instruction following, and fine-grained tags to control emotion, tone, persona, speaking rate, and volume. For voice cloning, it is more robust to noise and reverberation in reference audio, improving audio quality, clarity, and expressiveness. The Plus version emphasizes audio quality and resolution for professional scenarios such as brand voices, audio content creation, film and television dubbing, and character performance.
xiaomi/mimo-v2.6-pro-ultraspeed
MiMo-V2.6-Pro-UltraSpeed is the high-speed version of Xiaomi's model MiMo-V2.6-Pro. Built on the same 1-trillion-parameter MiMo-V2.6-Pro checkpoint, it delivers about 10 times faster output while maintaining the same quality as the original model. It supports a 1 million token context window and native multimodal input, including text, images, videos, and audio. Optimized for agent workflows, it delivers top-tier performance in coding, vision, general Q&A, and scientific research, and excels at complex long-horizon tasks.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen-image-2.1-turbo
Qwen-Image-2.1-Turbo is an open-source image generation model in the Qwen-Image series with excellent overall performance and cost-effectiveness. Qwen-Image-2.1 offers four key strengths: Compact Model, Strong Performance, Exceptional Value. Native Transparency, Unified Generation and Editing. Versatile Editing: Multiple References, Local Control, Fidelity, and Comprehensive Capabilities. Realistic Textures, Elegant Typography, and Beautiful Portraits.
qwen-image-2.1-pro
Qwen-Image-2.1-Pro is an image model in the Qwen family that balances generation quality, inference efficiency, and cost. Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model, with just 7B parameters in its visual generation component and native support for generating and editing transparent images. This update introduces four key improvements: Compact and efficient: A lightweight architecture and inference optimizations balance image quality and computational cost. Native transparency, unified creation and editing: Generate regular or transparent images from prompts, edit transparent layers, and extract subjects from photos. Versatile editing: Use up to 10 reference images, specify local edits, preserve people and products, and handle a wide range of tasks. Realistic textures and refined aesthetics: Improved typography, portrait lighting, and fine details make generated images more visually compelling.
decision-model-preview
A structured decision model designed for high-frequency business decisions. It performs classification, yes-or-no decisions, and scoring in parallel based on text or business context, returning probability distributions and confidence scores for scenarios such as ticket routing, content moderation, agent routing, and result validation.
stepfun/step-5-preview
Step 5 Preview is a foundation model for real-world Agentic tasks, achieving frontier performance in AI programming, software engineering, and professional knowledge work, with particularly outstanding capabilities in finance. It uses a sparse MoE architecture (600B total parameters, 27B active per token), supporting a 1M token context window and visual input.
qwen3.8-omni-flash-realtime
Supports real-time audio and video interaction with text and audio output, adding multichannel audio, video aggregation, and remote MCP tools. Supports WebSocket, WebRTC, and AOQ access.
ZHIPU/GLM-5.3-FlashX
GLM-5.3-FlashX is the high-speed model in the GLM-5 series, delivering inference speeds of up to 200 tokens/s for a faster, smoother model experience. It supports a 1M context and 128K output, and natively accepts image, video, and file input.
vanchin/deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the smallest model in DeepSeek's all-new model architecture series, with native multimodal visual understanding, a 1M context, and up to 384K output. The new architecture is designed for a higher capability ceiling, faster inference, greater throughput, and scalability to larger parameter models.
qwen-audio-3.1-tts-next
Qwen-Audio-3.1-TTS-Next is an AudioGen model designed for unified audio generation. Going beyond conventional TTS systems that focus solely on speech synthesis, it can generate complete audio content in a single pass from inputs such as text, timestamps, and reference audio, seamlessly combining speech, sound effects, ambient sounds, and more. The model supports a wide range of tasks, including multilingual TTS, single-speaker speech, multi-speaker dialogue, podcasts, cinematic soundscapes, ambient audio, and sound effects. Its key strengths include more natural and expressive speech, consistent voice identity, flexible timing control, and high-quality soundscape generation. These capabilities make it well suited for content creation, short-form video, podcasts, film and television production, game audio, and other professional audio creation scenarios.
qwen-audio-3.1-asr-flash-message
Qwen-Audio-3.1-ASR-Flash-Message is a generative speech recognition model optimized for voice messages, voice input, and enterprise scenarios. It supports Chinese, English, and other languages, with optimizations for Indonesian, Japanese, Thai, Filipino, Vietnamese, Korean, and Malay. The model supports incremental streaming transcription, P0/P1 hotwords, text normalization, recognition informed by preceding transcripts and business context, and optional native text polishing. It improves recognition stability with multiple speakers, background conversations, far-field audio, and environmental noise. It supports automatic language detection or specified target languages, and can transcribe Chinese dialects as spoken or convert them to standard Mandarin text.
qwen-audio-3.1-asr-flash-streaming
Qwen-Audio-3.1-ASR-Flash-Streaming is an end-to-end real-time speech recognition model designed for scenarios such as real-time conferencing, live subtitles, and intelligent interaction. It features low latency and high accuracy in streaming speech transcription. The model supports multilingual and multi-regional Chinese dialect recognition, seamless switching between Chinese and English, hot word and context enhancement, punctuation prediction, and text normalization. It also exhibits strong noise robustness and can adapt to complex acoustic environments. Furthermore, this version supports controllable ASR/AST output for multiple dialects.
qwen-audio-3.1-asr-flash
Qwen-Audio-3.1-ASR-Flash is a large-scale model supporting efficient short speech recognition, designed for high-quality, multilingual, and cultural content transcription scenarios. The model supports multilingual and multi-regional Chinese dialect recognition, and is optimized for the rhythm, cadence, and classical Chinese expression of ancient Chinese poetry. It features context enhancement, punctuation prediction, and text normalization capabilities. This version also supports controllable ASR/AST output for multiple dialects. It includes native transcription and polishing capabilities, as well as industrial-grade multi-person dialogue role-based transcription capabilities.
qwen-audio-3.1-asr-flash-filetrans
Qwen-Audio-3.1-ASR-Flash-Filetrans is an end-to-end offline speech recognition model designed for scenarios such as meeting transcription, content production, and call analysis. It supports multilingual and multi-regional Chinese dialect recognition. The model features high-precision transcription, hot word and context enhancement, speaker separation, punctuation prediction, and text normalization capabilities. It maintains stable performance even in complex noisy environments, making it the preferred choice for long audio transcription. This version also supports controllable ASR/AST output for multiple dialects.
glm-5.3-prime
GLM-5.3-Prime is a high-speed version of GLM-5.3. It supports a 1M-token context window, up to 128K output tokens, and three thinking-depth levels. It provides coding, agent, and cybersecurity capabilities, with a 50% improvement over its predecessor on an internal coding benchmark, and is suited to end-to-end project delivery and long-running complex tasks.
qwen-audio-3.1-realtime-plus
Supports real-time duplex speech conversations using the same integration protocol as 3.0 Plus. Retains existing voices and adds eight system voices, with a 262,144-token context window, function calling, web search, and voice cloning.
qwen-audio-3.1-tts-flash
Qwen-Audio-3.1-TTS-Flash is a high-performance text-to-speech model designed for real-time interactions. It supports multiple languages and dialects, streaming speech synthesis, free-style instruction following, and fine-grained tag control for emotion, tone, persona, speaking rate, and volume. In voice cloning scenarios, it is more robust to noise and reverberation in reference audio, with improved audio quality, clarity, and expressiveness. The Flash version is optimized for low-latency applications such as voice assistants, real-time conversations, and customer service.
happyoyster-1.0-adventure
HappyOyster Adventure is a real-time interactive, explorable open-world model. It generates world scenes from multimodal (text/image) input, enabling immersive, boundless exploration through real-time interaction, movement, and camera control.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
qwen-mt-uni
Qwen-MT-Uni translates text, documents (PDF/Word/PPT/Excel/HTML/Markdown/TXT), images, and audio in a single call. The server automatically detects the input modality and orchestrates layout restoration, voice cloning, and other post-processing, returning either the translated text or a signed URL to the translated file. Supports both synchronous and asynchronous invocation for large-scale cross-modal translation scenarios.
kimi/kimi-k2.8-preview
Kimi K2.8 Preview is served directly by Moonshot AI. It offers overall performance close to K3 with improved thinking efficiency. It supports the same thinking-effort levels as K3 and a one-million-token context window.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-livetranslate-flash-realtime
The real-time version of Qwen3.8-LiveTranslate-Flash provides multilingual simultaneous translation for live audio and video. Built on Qwen-Omni with cross-language and cross-modal alignment and visual enhancement, it understands 60 languages and speaks 29 languages.
qwen3.8-max-0902
Qwen3.8-Max-0902(alias qwen3.8-max-2026-09-02)is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
qwen3.7-text-rerank
Based on Qwen3.7, this multilingual text reranking model evaluates the relevance between input queries and candidate documents. Compared to qwen3-rerank, it demonstrates significant improvements in general and web search, code retrieval, instruction following, as well as agentic and memory retrieval. It achieves superior performance across evaluation tasks such as MTEB and web search, delivering an approximate 35% relative improvement in web search.
qwen3.7-text-embedding-flash
A Flash multilingual text embedding model based on Qwen3.7, designed for cost- and throughput-sensitive scenarios. Compared to text-embedding-v4, it features significant upgrades in multilingual coverage (from 100 to 201 languages), long-text processing (from 32K to 128K), and Sparse Embedding capabilities. It maintains comparable overall performance in semantic, cross-lingual, and code retrieval evaluations, while achieving an approximately 3% improvement on the AIRBench-zh benchmark.
ZHIPU/GLM-5.3-Flash
GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. With 320B total parameters and 18B activated parameters, it combines sparse attention with linear attention to deliver intelligence beyond GLM-5.2 at a lower cost. It supports a 1M context window and 128K output, natively accepts image, video, and file input, and can work across code, browsers, and graphical interfaces.
qwen-mt-image-2.0
An advanced image translation engine supporting 55 source languages. It delivers pixel-perfect replication of the original layout and typography. Gain full control over professional content with custom terminology and sensitive word filtering, ensuring a flexible, accurate, and high-performance image localization service.
glm-5.3
GLM-5.3 is Z.ai's large language model, delivering enhanced coding and agentic capabilities, with a 50% improvement over its predecessor on Z.ai Code Bench and end-to-end delivery of production-grade results. It supports cybersecurity tasks such as code vulnerability discovery and delivers excellent performance on the CyberGym benchmark. With a 1M-token context window, 128K-token maximum output, and three reasoning effort levels, GLM-5.3 excels at long-horizon, complex tasks.
kling/kling-v3-turbo-video-generation
Kling V3 Turbo ultra-fast video generation model with built-in audio-video synchronization. Supports text-to-video, image-to-video (first frame), and multi-shot generation.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
vidu/vidu-image-pro_reference2image
Input 0-14 reference images or text descriptions. Supports reference-based image generation, text-to-image, and image editing with precise rendering of Chinese and English text, pixel-level restoration of UI/chart design details. Suitable for creating posters, infographics, etc.
vidu/vidu-image-lite_reference2image
Input 0-14 reference images or text descriptions. Supports reference-based image generation, text-to-image, and image editing with precise rendering of Chinese and English text, pixel-level restoration of UI/chart design details. Suitable for creating posters, infographics, etc.
vanchin/deepseek-v4-pro-0813
DeepSeek-V4-Pro-0813 is a mixture-of-experts model served directly by Vanchin. It has 1.6T total parameters, activates 49B parameters per token, and supports a one-million-token context window. It was pretrained on more than 32T high-quality, diverse tokens.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
qwen3.8-27b
A 27B native vision-language Dense model of the Qwen3.8 series. Compared with 3.6-27B, it focuses on improving coding and office-scenario capabilities under text and visual modalities, able to more reliably complete complex tasks end-to-end and deliver trustworthy results.
ZHIPU/GLM-5.3
GLM-5.3 is Zhipu's coding model, improving 50% over GLM-5.2 in internal perceptual evaluations. In cybersecurity, GLM-5.3 matches Mythos 5 in white-box code review and vulnerability discovery, demonstrating strong potential for cybersecurity defense scenarios.
MiniMax/MiniMax-H3
MiniMax H3 is a general-purpose omni-modal generation system. It can uniformly understand multimodal context composed of text, images, video, and audio, and can generate videos up to 2K resolution, 15 seconds long, with native stereo audio. Thanks to a task-generalization-oriented system design, H3 already possesses broad multimodal context understanding and generation capabilities at the pretraining stage, enabling it to excellently follow complex multimodal instructions.
unisound/unisound-u2
Unisound U2 is a native agent model served directly by Unisound. It supports long-text comprehension, task decomposition, tool calling, and multistep execution. It is suited to software development, complex reasoning, in-depth research, enterprise analysis, and automated workflows.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
qwen3.8-2.4t-a95b
Qwen3.8-2.4T-A95B is the open-source version of Qwen's series, released in August 2026. It adopts a sparse MoE architecture with 2.4 trillion total parameters and about 95 billion activated per step, combined with a hybrid attention mechanism, supporting a 1 million token context. Key benchmarks: GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, BabyVision 82.0, CodeArena global #4.
pixverse/pixverse-v6-r2v-omni
V6-r2v-omni is a new model released by PixVerse in August 2026. It adds a new capability on top of the existing Video (T2V/I2V), Transition (first and last frames), and Extend features, supporting mixed video and image input to generate videos.
wan3.0-video
Wan3.0 is an all-in-one video generation model that uniformly supports multiple creative functions including reference, editing, replication, and driving. It supports four-modal all-reference input, up to 30 seconds of video generation, and can parse files, web pages, and complex images. With production-grade character consistency and realistic audio and visuals, it delivers an immersive audio-visual impact.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
qwen-audio-3.0-asr-flash-streaming
qwen-audio-3.0-asr-flash-filetrans
qwen-audio-3.0-asr-flash
Added the Qwen-Audio-3.0-ASR-Flash-Streaming (real-time), Qwen-Audio-3.0-ASR-Flash-Filetrans (non-real-time), and Qwen-Audio-3.0-ASR-Flash (non-real-time) models: Dialect support: Supports the seven major Chinese dialect groups (Mandarin, Wu, Xiang, Gan, Hakka, Min, and Yue) and more than 20 regional accents; Classical poetry optimization: Improves recognition accuracy for classical Chinese poetry, making it suitable for education, culture, and audiobook scenarios; Text optimization: Enhances punctuation prediction and text normalization, automatically converting numbers, dates, and monetary amounts to standard formats; Multilingual expansion: Supports 30 languages, including Chinese, English, Japanese, and Korean; Hotwords and context: Supports hotwords (precompiled and on-the-fly) and context input to improve recognition accuracy for domain-specific terms.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
kimi/kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
qwen3.7-text-embedding
Qwen3-7-Text-Embedding is a multilingual text vector model based on Qwen3.7, offering significant improvements in text retrieval, clustering, and classification over text-embedding-v4. It achieves 20% better performance in MTEB multilingual, Chinese-English, and code retrieval tasks and supports customizable vector dimensions (256-2560).
qwen-audio-3.0-tts-plus
Qwen-Audio-3.0-TTS-Plus is a high-performance speech synthesis model, designed for high-quality speech generation scenarios. Compared with the previous version, it supports more low-resource languages and Chinese dialects, significantly improves dialect authenticity, and enhances free-style instruction following and fine-grained tag control for more accurate control over emotion, tone, character, speaking rate, volume, and synthesis style. It is also more robust under noisy and reverberant acoustic conditions, with further improvements in audio quality, clarity, resolution, and overall expressiveness. The Plus version focuses more on synthesis quality and detailed expressiveness, making it suitable for professional scenarios with higher requirements for audio quality, naturalness, and expressiveness, such as content creation, audiobooks, film and video dubbing, brand voice design, and premium speech services.
qwen-audio-3.0-tts-flash
qwen-audio-3.0-tts-flash is a high-performance speech synthesis model, optimized for real-time interactive scenarios. Compared with the previous version, it supports more low-resource languages and Chinese dialects, improves dialect authenticity, and enhances free-style instruction following and fine-grained tag control for more flexible control over emotion, tone, character, speaking rate, volume, and expressive style. It is also more robust under noisy and reverberant acoustic conditions, with improved audio quality, clarity, and overall expressiveness. The Flash version focuses on real-time synthesis, with first-packet latency controlled within 200 ms, making it suitable for voice assistants, real-time dialogue, intelligent customer service, and other low-latency interactive applications.
qwen-audio-3.0-realtime-plus
Qwen 3.0 Realtime Speech Model Standard Edition - next-generation duplex speech model ranked #1 globally in Artificial Analysis Speech-to-Speech benchmark. Balances model intelligence with duplex dialogue rhythm for natural interaction and enhanced response quality.
qwen-audio-3.0-realtime-flash
Qwen 3.0 Realtime Speech Dialogue Model Flash Edition - a next-generation duplex speech model ranked #1 globally in Artificial Analysis Speech-to-Speech benchmark. Combines high intelligence with optimized duplex rhythm for low-latency (parallel inference, full-streaming optimization) 'fast and smart' dialogue experience focusing on response speed.
pixverse/pixverse-upscale
Upscale enhances videos of any resolution to 4K, improving clarity and detail for high-quality presentation and redistribution.
pixverse/pixverse-motioncontrol
Motion Control extracts actions from reference videos and transfers them to target character images, generating new videos of characters performing the same actions.
pixverse/pixverse-lipsync
Lip-sync capability aligns characters' mouth movements in videos with input audio/TTS, enhancing natural speech and narrative expressiveness.
vidu/viduq3-pro-fast_img2video
Accepts images and text prompts for video generation. ViduQ3-Pro-fast offers faster generation speed and higher cost-effectiveness. Compared to ViduQ2-Pro-fast, it extends generation duration from 10s to 16s, enabling more complex shot transitions and narrative logic.
vidu/viduq3-fast_reference2image
Accepts 0-14 reference images or text prompts for reference-to-image generation, text-to-image generation, and image editing. Prioritizes high speed, quality, and low cost, with costs approximately 50% lower than Pro.
vidu/viduq3-drama_reference2video
ViduQ3-Drama is a premium drama/AI manga production model, upgraded across 'consistency, motion effects, detail aesthetics, and cost-effectiveness'. Delivers stable visuals, authentic emotions, and enhanced dynamics.
vidu/viduq3-ad_reference2video
ViduQ3-Ad is an advertising-specific model featuring 'marketing-grade scene transitions, intelligent camera movement, and direct audio output'. Upload a product image to generate a 16-second ad video, lowering production barriers and costs.
vidu/viduq2-pro_reference2image
Generates images from 0-14 references or text, excelling in complex logic with strong context consistency and industrial-grade stability. Suitable for professional design and web drama production.
vidu/viduq2-fast_reference2image
Generates images from 0-14 references or text, with enhanced semantic understanding and support for diverse styles.
vidu/vidu-image_reference2image
Generates images from 0-14 reference images or text, supporting reference-based generation, text-to-image, and editing. Ensures pixel-level accuracy for Chinese/English text and UI/chart details, ideal for posters and infographics.
glm-5.2-fast-preview
GLM-5.2-Fast-Preview is the high-speed variant of Zhipu AI's GLM-5.2, with 1M context and capabilities on par with the standard version. Inference-optimized to deliver 1.5–2× the output TPS, it fits latency-sensitive use cases such as real-time chat, multi-turn agents, and streaming code generation.
wan2.7-t2v-2026-06-12
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of June 12, 2026.
wan2.7-r2v-2026-06-12
Wan2.7 reference to video, enhanced consistency & performance. Delivering superior stability for characters, props, and scenes. Supports hybrid referencing of up to 5 mixed image/video inputs and audio timbre cloning. Together with core engine upgrades, it achieves unprecedented cinematic expressive power. This version is a snapshot as of June 12, 2026.
qwen-image-2.0-pro-2026-06-22
The Qwen-Image-2.0 series full-fledged model integrates image generation and editing; it boasts more professional text rendering capabilities with 1k token command support, more delicate and realistic textures, meticulous depiction of realistic scenes, and stronger semantic adherence. The full-fledged version possesses text rendering capabilities and realistic textures in the 2.0 series.
kimi/kimi-k2.7-code-highspeed
K2.7 Code High-Speed Edition shares the same model as the standard version but delivers 5-6x faster output speed, achieving ~180 tokens/s in median programming scenarios and ~260 tokens/s in short-context scenarios, enhancing programming efficiency.
fun-asr-flash-2026-06-15
The Bailing ASR version, updated in June 2026, fully supports the seven major dialect systems of traditional Chinese (Mandarin/Wu/Xiang/Gan/Hakka/Min/Yue) and is compatible with over 20 regional Mandarin accents. It features specific optimizations for the rhythm, meter, and literary expression characteristics of classical Chinese poetry, improving the accuracy of poetry recognition and making it suitable for cultural heritage, educational explanations, and audiobooks. Improved punctuation prediction and text normalization capabilities make the output text more consistent with written expression habits; numbers, dates, and amounts are automatically converted to standard formats, enhancing readability and professionalism. The language support has been expanded to include English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, and Slovak, totaling 30 languages. It supports contextualization capabilities and can transcribe audio up to 5 minutes in length.
qwen3.5-ocr
Qwen3.5-OCR is an upgraded OCR model with enhanced document parsing, text localization, and key information extraction. It significantly improves extraction performance for real-world documents (e.g., ID cards, driver's licenses).
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
glm-5.2
ZHIPU/GLM-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
kimi-k2.7-code
kimi/kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
qwen3.7-max-2026-06-08
The Max model, the largest in the Qwen3.7 series, has added visual‑modal understanding compared to the May 20 snapshot, enabling it to perceive real‑world scenes and supporting multimodal interactive hybrid agent capabilities. This version is based on a snapshot taken on June 8, 2026.
pixverse/pixverse-v6-r2v
V6 is a new model launched by PixVerse in late March 2026. The r2v (multi-subject reference video generation) model ranks second globally, accepting 2-7 input images to intelligently fuse different subjects. It excels in complex mid-shot and long-shot video scenarios while retaining t2v's prompt control capabilities and it2v's consistency preservation. It offers enhanced emotional expression and smoother high-speed motion visuals, supports 15-second long videos, direct music-and-video output, and multiple languages. It enables one-click generation for scenarios like e-commerce product close-ups, advertisement trailers, and simulating C4D modeling to showcase product structures.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
fun-music-preview
The Fun Music large model supports song creation based on open-ended lyrics or composition requirements, generating full Chinese/English songs with male/female vocals. The songs are accessible and emotionally progressive, representing a perfect fusion of human inspiration and LLM capabilities.
MiniMax/MiniMax-M3
MiniMax M3 excels in enterprise-grade long-document understanding, high-quality content generation, code writing, bug fixing, and native application building, powered by industry-leading coding and agentic capabilities, a 1M ultra-long context window, and native multimodal features. Its agentic capabilities enable end-to-end workflow integration, while native multimodal support delivers seamless text-image interaction.
vanchin/deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
stepfun/step-3.7-flash
Step 3.7 Flash is a production-grade Agent-efficient Flash model newly launched by Stepfun for building high-performance Agents. Optimized for speed, cost efficiency, execution reliability, and complex task completion, it features multimodal perception, visual search, tool enhancement, high-reliability tool orchestration, and Agent ecosystem compatibility.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
qwen3.7-max-preview
qwen3.7-max-2026-05-17
The Max model, the largest in the Qwen3.7 series, is available as a preview and supports only thinking mode, offering a pure text‑only interface for experimentation. It is primarily optimized for general‑purpose conversational use cases, such as knowledge‑based question answering, instruction following, and creative writing.
qwen3.5-livetranslate-flash-realtime
qwen3.5-livetranslate-flash-realtime-2026-05-19
The real-time version of Qwen3.5-LiveTranslate-Flash, which is a high-precision, highly responsive, and robust multilingual simultaneous audio and video interpretation model. Leveraging Qwen3.5-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3.5-LiveTranslate-Flash offers both offline and real-time audio and video translation capabilities. It can understand 60 languages and speak 29 languages.
xiaomi/mimo-v2.5-pro
MiMo-V2.5-Pro is Xiaomi's model. Compared to predecessors, it significantly improves general agent capabilities, complex software engineering, and long-horizon tasks, leading benchmarks like ClawEval, GDPVal, and SWE-bench Pro. It autonomously completes expert-level tasks requiring thousands of tool calls over days/weeks, with a 1 million token context length ideal for agent frameworks.
ZHIPU/GLM-5.1
GLM-5.1 is Zhipu's model with greatly enhanced coding ability and significantly improved long-horizon task performance. It can work continuously and autonomously for up to 8 hours in a single task, completing the full closed loop from planning and execution to iterative optimization and delivering engineering-grade results. In overall and coding capabilities, GLM-5.1 aligns with Claude Opus 4.6, and demonstrates stronger sustained working ability in long-horizon autonomous execution, complex engineering optimization, and real-world development scenarios, making it an ideal foundation for building Autonomous Agents and long-horizon Coding Agents.
fun-music-v1
A music generation model creating complete Chinese/English songs (male/female vocals) from lyrics or creative requirements, blending human inspiration with AI capabilities.
vidu/viduq3-mix_reference2video
Accepts 1-7 reference images and text prompts for video generation. ViduQ3-mix delivers strong visual quality and balanced performance. Designed for episodic content creation, it integrates 6 visual effects (particle/fluid/dynamics/camera movement/transition/lighting), 5 audio effects (environment/dynamic/atmosphere/sound design/emotional), and 4 scenarios (short films/animated series/TV dramas/advertisements), enabling versatile creation of short films, animations, and ads.
vidu/viduq3_reference2video
Accepts 1-7 reference images and text prompts for video generation. ViduQ3 supports intelligent scene transitions with enhanced multi-camera consistency. Designed for episodic content creation, it integrates 6 visual effects (particle/fluid/dynamics/camera movement/transition/lighting), 5 audio effects (environment/dynamic/atmosphere/sound design/emotional), and 4 scenarios (short films/animated series/TV dramas/advertisements), enabling versatile creation of short films, animations, and ads.
vidu/viduq3-turbo_reference2video
Accepts 1-7 reference images and text prompts for video generation. ViduQ3-Turbo offers fast generation speed and high cost-effectiveness. Designed for episodic content creation, it integrates 6 visual effects (particle/fluid/dynamics/camera movement/transition/lighting), 5 audio effects (environment/dynamic/atmosphere/sound design/emotional), and 4 scenarios (short films/animated series/TV dramas/advertisements), enabling versatile creation of short films, animations, and ads.
vidu/viduq2-pro-fast_img2video
Generates videos from images and text. ViduQ2-Pro-fast offers cost-effective, stable performance with 2-3x faster speed than turbo. As the world's first 'Everything as Reference' video model, it supports reference in six dimensions (effects, expressions, textures, actions, characters, scenes) for precise editing, tailored for web dramas, short films, and production.
Tripo/Tripo-P1.0
Tripo P1.0 is a real-time 3D generation model for developers and creators needing clean topology and engine-ready meshes. It generates professional-grade 3D assets in ~2 seconds, optimized for gaming, Web3D, and interactive scenarios. It prioritizes speed and out-of-the-box usability for UGC pipelines, accelerating integration into real-time engines.
Tripo/Tripo-H3.1
Tripo H3.1 is a high-precision 3D generation model designed for creators requiring extreme visual quality and detail. With 20B-level parameters, it supports billion-voxel resolution and up to 2M-polygon generation. It enhances geometric fidelity and texture alignment for complex structures like character forms, facial details, and text, suitable for high-quality visual production and 3D printing.
wan2.7-t2v-2026-04-25
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of April 25, 2026.
wan2.7-i2v-2026-04-25
Wan2.7 image to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of April 25, 2026.
kimi/kimi-k2.6
Kimi K2.6 is a model in the Kimi series, featuring comprehensive enhancements in general-purpose agent capabilities, coding, and visual understanding. It achieves industry-leading results across multiple benchmarks, including the PhD-level full version of Humanity's Last Exam, SWE-Bench Pro for real-world software engineering, and DeepSearchQA for agent deep retrieval. The model supports text, image, and video inputs, offers both thinking and non-thinking modes, and handles conversational and agent tasks.
happyhorse-1.0-video-edit
HappyHorse-1.0-V2V supports advanced video editing through natural language instructions. It allows for local or global editing of video elements using up to 5 reference images, precisely preserving original motion dynamics to achieve superior expressiveness.
happyhorse-1.0-r2v
HappyHorse-1.0-R2V supports Reference-to-Video generation, offering enhanced stability in subject and scene referencing. Capable of processing up to 9 reference images, it precisely preserves creative intent to deliver superior performance.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
qwen3.5-plus-2026-04-20
The Qwen3.5 native vision-language series Plus model has seen a substantial improvement in agentic coding capabilities compared to the February 15th snapshot. Inference speed has also been significantly enhanced, while its knowledge retention, reasoning ability, and long-context processing remain at a high level, making it well-suited for complex agent-based tasks. It is ideal for applications such as coding agents, production workflows, and high-throughput scenarios. This version is based on a snapshot taken on April 20, 2026.
qwen-image-2.0-pro
qwen-image-2.0-pro-2026-04-22
The full-featured Qwen-Image-2.0 series models integrate image generation and image editing, offering enhanced text rendering with support for 1,000-token prompts, more refined realistic textures, detailed depiction of photorealistic scenes, and stronger semantic adherence. The full-featured version delivers text rendering and realistic textures in the 2.0 series.
qwen3.6-27b
The Qwen3.6 27B native vision-language dense model builds upon the 3.5-27B version, with key improvements in agentic coding capabilities and enhanced STEM reasoning and inference skills. In the vision modality, it demonstrates significant advances in spatial intelligence, object localization, and detection, while video understanding, document OCR, and visual agent capabilities continue to improve steadily.
happyhorse-1.0-i2v
HappyHorse-1.0-I2V enables image-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends both text and image semantics to produce high-quality videos that are fluid, natural, and rich in detail.
kimi-k2.6
Kimi-k2.6 is an intelligent model in Kimi series, featuring enhanced long-context code generation, improved instruction following and self-correction capabilities, supporting text/image/video inputs and multiple operational modes.
happyhorse-1.0-t2v
HappyHorse-1.0-T2V supports text-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends text semantics to produce high-quality videos that are fluid, natural, and rich in detail.
qwen3.6-max-preview
The Max model, the largest in the Qwen3.6 series, is now available in a preview version. At present, only its plain-text capabilities are open for experimentation. Compared with the previously released Qwen3-Max and Qwen3.6-Plus, this model features enhanced vibe coding abilities, more efficient coding agent execution, and significantly improved front-end development skills. Additionally, its long-tail knowledge retention has been further upgraded.
qwen3.6-flash
qwen3.6-flash-2026-04-16
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen3.6-35b-a3b
The Qwen3.6 35B-A3B native vision-language model is built on a hybrid architecture that integrates linear attention mechanisms with a sparse mixture-of-experts framework, achieving higher inference efficiency. Compared with the 3.5-35B-A3B, this model demonstrates significantly improved agentic coding capabilities, mathematical and code reasoning abilities, spatial intelligence, as well as object localization and object detection performance.
glm-5.1
GLM-5.1 is a model developed by Zhipu AI, specifically designed for long-horizon tasks. It has 744 billion parameters, supports an ultra-long context of 200k tokens, and can generate up to 128k tokens in a single response. GLM-5.1 excels in logical reasoning, long-text understanding, and code generation, while balancing performance with inference efficiency. It delivers outstanding results across multiple multi-task benchmarks and is well-suited for applications such as intelligent human-computer interaction, enterprise solutions, and developer assistance.
pixverse/pixverse-c1-t2v
C1, PixVerse's film-industry large model launched in March 2026, uses t2v (text-to-video) to precisely control visuals via prompts, accurately reproducing cinematic techniques (push-ins, pull-outs, pans, tracking shots). Supports 15s videos, direct music/video output, and multilingual text.
pixverse/pixverse-c1-r2v
C1, PixVerse's film-industry large model launched in March 2026, employs r2v (multi-reference-to-video) to input 2-7 images, intelligently fusing multiple subjects. Combines t2v prompt control with it2v consistency and cinematic combat/action effects. Supports 15s videos, direct music/video output, and multilingual text. Ideal for multi-character scenes, dialogues, and interactions with medium/long shots. If a multi-panel storyboard (up to 3x3 grid) is provided, generates continuous storyboard videos with one click.
pixverse/pixverse-c1-kf2v
C1, PixVerse's film-industry large model launched in March 2026, uses kf2v (keyframe-to-video) to seamlessly connect any two images with natural transitions. Supports 15s videos, direct music/video output, and multilingual text.
pixverse/pixverse-c1-it2v
C1, PixVerse's film-industry large model launched in March 2026, extends t2v prompt control with it2v (image-to-video) capabilities to precisely replicate reference images' colors, saturation, scenes, and character features. Compared to V6, it enhances prompt expressiveness, imagination, cinematic combat actions, and visual effects. Supports 15s videos, direct music/video output, and multilingual text. Suitable for close-ups, monologues, stop-motion/slow-motion, and empty-shot transitions.
vanchin/deepseek-v3.2-think
DeepSeek-V3.2 harmonizes high computational efficiency with superior reasoning and agent performance. Built on DeepSeek-V3, it incorporates key innovations like DeepSeek Sparse Attention (DSA), scalable reinforcement learning frameworks, and large-scale agent task synthesis pipelines, advancing the frontier of open-source LLMs.
vanchin/deepseek-v3.1-terminus
DeepSeek-V3.1-Terminus is an updated version of DeepSeek-V3.1, maintaining the model's core capabilities while addressing user feedback through fixes and optimizations. It retains the same architecture as DeepSeek-V3 and achieves significant enhancements in specific domains.
vanchin/deepseek-v3
Released by DeepSeek in December 2024, DeepSeek-V3 is a leading open-source Mixture-of-Experts (MoE) language model with 671B total parameters, activating only 37B parameters per token. Pre-trained on 14.8 trillion high-quality tokens, it natively supports a 128K context window. The model leverages an innovative auxiliary-loss-free load balancing strategy, a Multi-head Latent Attention (MLA) architecture, and FP8 mixed-precision training.
vanchin/deepseek-r1
DeepSeek-R1 is a 671B-parameter Mixture of Experts (MoE) reasoning model open-sourced by DeepSeek in January 2025, activating only 37B parameters during inference. It is the first 100B+ parameter model trained purely via reinforcement learning without supervised fine-tuning, enabling the natural emergence of Chain-of-Thought (CoT). By incorporating cold-start data prior to RL, the model resolves the repetition and language mixing issues of R1-Zero, achieving performance on par with OpenAI o1 across math, coding, and reasoning tasks.
vanchin/deepseek-ocr
DeepSeek-OCR, centered on 'exploring vision-text compression boundaries', redefines the functional positioning of vision encoders from a large language model (LLM) perspective, providing a new solution balancing accuracy and efficiency for document recognition, image-to-text conversion, and other high-frequency scenarios.
wan2.7-videoedit
Wan2.7 video edit, supports both localized and global editing with prompt. Seamlessly replace elements using image references and replicate complex dynamic processes, including motion, special effects, and camera movements.
wan2.7-t2v
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-r2v
Wan2.7 reference to video, enhanced consistency & performance. Delivering superior stability for characters, props, and scenes. Supports hybrid referencing of up to 5 mixed image/video inputs and audio timbre cloning. Together with core engine upgrades, it achieves unprecedented cinematic expressive power.
wan2.7-i2v
Wan2.7 image to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-image-pro
Wan2.7–image-pro, supports text to image, text/image to sequential images, image editing, multi-image reference generation, and interactive editing. Delivers enhanced performance in text rendering, subject consistency, and complex instruction following.
wan2.7-image
Wan2.7 – image generation and editing, supports text to image, text/image to sequential images, image editing, multi-image reference generation, and interactive editing. Delivers enhanced performance in text rendering, subject consistency, and complex instruction following.
qwen3.6-plus
qwen3.6-plus-2026-04-02
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
pixverse/pixverse-v6-t2v
V6 is a new model launched by PixVerse in late March 2026. The t2v (text-to-video) model enables precise video control through prompts, accurately reproducing various cinematic techniques including push-in, pull-out, panning, tracking, and follow shots with natural transitions and precise perspective control. It supports 15-second long videos, direct music-and-video output, and multiple languages.
pixverse/pixverse-v6-kf2v
V6 is a new model launched by PixVerse in late March 2026. The kf2v (keyframe-to-video) model seamlessly connects any two images, achieving smooth and natural video transitions. It supports 15-second long videos, direct music-and-video output, and multiple languages.
pixverse/pixverse-v6-it2v
V6 is a new model launched by PixVerse in late March 2026. The it2v (image-to-video) model ranks second globally, offering not only the prompt control capabilities of t2v (text-to-video) but also high-fidelity reproduction of reference images' colors, saturation, scenes, and character features. It provides enhanced emotional expression and high-speed motion performance, supports 15-second long videos, direct music-and-video output, and multiple languages. It enables one-click generation for scenarios like e-commerce product close-ups, advertisement trailers, and simulating C4D modeling to showcase product structures.
qwen3.5-omni-plus-realtime
qwen3.5-omni-plus-realtime-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a fully evolved version of Qwen3-Omni, it supports audio input in 60+ languages, voice output in 30+ languages, and controllable voice dialogue, WebSearch and complex FunctionCall invocation, and has intelligent semantic interruption interaction capabilities. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interactive experience.
qwen3.5-omni-plus
qwen3.5-omni-plus-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a comprehensive evolution of Qwen 3-Omni, it supports over 10 hours of audio understanding and over 400 seconds of 720P (1 FPS) audio-visual understanding and dialogue. It further expands the language range, supporting audio input in 60+ languages and speech output in 30+ languages. It also possesses powerful structured audio-visual understanding capabilities and is widely used in text creation, voice assistants, multimedia analysis, and other scenarios, providing a natural and fluent multimodal understanding and interactive experience.
qwen3.5-omni-flash-realtime
qwen3.5-omni-flash-realtime-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a fully evolved version of Qwen3-Omni, it supports audio input in 60+ languages, voice output in 30+ languages, and controllable voice dialogue, WebSearch and complex FunctionCall invocation, and has intelligent semantic interruption interaction capabilities. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interactive experience.
qwen3.5-omni-flash
qwen3.5-omni-flash-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a comprehensive evolution of Qwen 3-Omni, it supports over 10 hours of audio understanding and over 400 seconds of 720P (1 FPS) audio-visual understanding and dialogue. It further expands the language range, supporting audio input in 60+ languages and speech output in 30+ languages. It also possesses powerful structured audio-visual understanding capabilities and is widely used in text creation, voice assistants, multimedia analysis, and other scenarios, providing a natural and fluent multimodal understanding and interactive experience.
vidu/viduq3-turbo_text2video
Accepts text prompts for video generation. ViduQ3-Turbo is a high-performance accelerated model with extreme generation efficiency, combining superior image quality and dynamic performance. Excels in combat scenes, emotional rendering, and semantic understanding. Outstanding cost-effectiveness for image-based social platforms, AI companionship, and VFX material creation.
vidu/viduq3-turbo_start-end2video
Accepts start/end frames and text prompts for video generation. ViduQ3-Turbo is a high-performance accelerated model with extreme generation efficiency, combining superior image quality and dynamic performance. Excels in combat scenes, emotional rendering, and semantic understanding. Outstanding cost-effectiveness for image-based social platforms, AI companionship, and VFX material creation.
vidu/viduq3-turbo_img2video
Accepts images and text prompts for video generation. ViduQ3-Turbo is a high-performance accelerated model with extreme generation efficiency, combining superior image quality and dynamic performance. Excels in combat scenes, emotional rendering, and semantic understanding. Outstanding cost-effectiveness for image-based social platforms, AI companionship, and VFX material creation.
vidu/viduq3-pro_text2video
Accepts text prompts for video generation. ViduQ3-Pro is a audiovisual-native model supporting up to 16s of synchronized audio-visual generation. It enables free multi-shot transitions with precise control over pacing, emotions, and narrative continuity. Leading parameter scale ensures cinematic-grade quality with superior image fidelity, character consistency, and emotional expression, suitable for professional production in advertising (e-commerce, TVC, performance ads), animated series, live-action dramas, and gaming.
vidu/viduq3-pro_start-end2video
Accepts start/end frames and text prompts for video generation. ViduQ3-Pro is a audiovisual-native model supporting up to 16s of synchronized audio-visual generation. It enables free multi-shot transitions with precise control over pacing, emotions, and narrative continuity. Leading parameter scale ensures cinematic-grade quality with superior image fidelity, character consistency, and emotional expression, suitable for professional production in advertising (e-commerce, TVC, performance ads), animated series, live-action dramas, and gaming.
vidu/viduq3-pro_img2video
Accepts images and text prompts for video generation. ViduQ3-Pro is a audiovisual-native model supporting up to 16s of synchronized audio-visual generation. It enables free multi-shot transitions with precise control over pacing, emotions, and narrative continuity. Leading parameter scale ensures cinematic-grade quality with superior image fidelity, character consistency, and emotional expression, suitable for professional production in advertising (e-commerce, TVC, performance ads), animated series, live-action dramas, and gaming.
vidu/viduq2_text2video
Generates videos from text. ViduQ2-text2video ensures precise instruction following and emotional nuance capture, with strong narrative control, micro-expression rendering, and dynamic cinematography. Widely used in film, advertising, web dramas, and tourism.
vidu/viduq2_reference2video
Generates videos from reference images and text. ViduQ2-reference2video ensures precise instruction following and emotional nuance capture, with strong narrative control, micro-expression rendering, and dynamic cinematography. Widely used in film, advertising, web dramas, and tourism.
vidu/viduq2-turbo_start-end2video
Generates videos from start/end frames and text. ViduQ2-Turbo is an ultra-fast engine: 720P 5s video in 19s, 1080P in 27s. Delivers natural human actions/expressions and high-quality motion effects for dynamic scenes like combat.
vidu/viduq2-turbo_img2video
Generates videos from images and text. ViduQ2-Turbo is an ultra-fast engine: 720P 5s video in 19s, 1080P in 27s. Delivers natural human actions/expressions and high-quality motion effects for dynamic scenes like combat.
vidu/viduq2-pro_start-end2video
Generates videos from start/end frames and text. ViduQ2-Pro is the world's first 'Everything as Reference' video model, supporting six-dimensional reference for precise editing. Designed for web dramas, short films, and production.
vidu/viduq2-pro_reference2video
Generates videos from reference videos, images, and text. ViduQ2-Pro-reference2video is the world's first 'Everything as Reference' video model, supporting six-dimensional reference for advanced editing. Enables controlled modifications for precise video creation, tailored for web dramas, short films, and production.
vidu/viduq2-pro_img2video
Generates videos from images and text. ViduQ2-Pro is the world's first 'Everything as Reference' video model, supporting six-dimensional reference for comprehensive editing. Enables controlled modifications for precise video creation, optimized for web dramas, short films, and production.
kling/kling-v3-video-generation
Intelligent scene segmentation interprets script transitions, automatically adjusting camera angles and shot types. Native multimodal framework ensures audio-visual coherence. Breaks duration limits for flexible multi-shot storytelling.
kling/kling-v3-omni-video-generation
Introduces 'All-in-One Reference' supporting 3-8s videos or multi-image anchoring for character elements. Synchronizes voice dubbing and lip movements for authentic character portrayal. Enhances video consistency and dynamic expression with audio-visual synchronization and intelligent scene segmentation.
kling/kling-v3-omni-image-generation
Unlocks cinematic narrative visuals with new series generation and 2K/4K direct output. Deeply analyzes audio-visual elements in prompts, accurately executing creative instructions. Supports multi-reference images and comprehensive effect upgrades, ideal for storyboarding, plot concept art, and scene design.
kling/kling-v3-image-generation
Supports up to 10 reference images, locking subjects, elements, and tones for style consistency. Integrates style transfer, portrait/character reference, multi-image fusion, and localized redrawing with flexible operations. Delivers realistic portrait details, refined visuals, and cinematic color atmospheres.
ZHIPU/GLM-5
GLM-5 targets coding and agent scenarios, achieving open-source SOTA in complex systems engineering with capabilities approaching Claude Opus, built on a 744B foundation with asynchronous reinforcement learning and sparse attention.
qwen-deep-research-2025-12-15
Qwen-Deep-Research is an advanced agent system for complex research tasks, equipped with multi-step reasoning and global planning capabilities. It leverages tools like internet search to perform detailed task decomposition, reasoning, and analysis, ultimately generating traceable, logically rigorous research reports.
tongyi-embedding-vision-plus-2026-03-06
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
tongyi-embedding-vision-flash-2026-03-06
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
MiniMax/MiniMax-M2.7
M2.7 autonomously constructs complex agent frameworks, leveraging capabilities like Agent Teams, complex Skills, and Tool Search to execute highly sophisticated productivity tasks.
pixverse/pixverse-v5.6-t2v
Input text descriptions to generate high-quality videos with precise semantic matching in seconds, supporting multiple styles. PixVerse V5.6 is a proprietary video generation large model developed by A-Soul Technology, featuring comprehensive upgrades in text-to-video and image-to-video capabilities. The model significantly improves image clarity, complex motion stability, and audio-visual synchronization, achieving more accurate lip-syncing and natural emotional expression in multi-character dialogue scenarios. It also optimizes composition, lighting, and texture consistency for further enhanced generation quality. PixVerse V5.6 ranks in the global top tier of the Artificial Analysis text-to-video and image-to-video leaderboards.
pixverse/pixverse-v5.6-r2v
Input 2-7 images to intelligently fuse different subjects while maintaining style consistency and motion coordination, enabling easy construction of rich narrative scenes with enhanced content controllability and creative flexibility. PixVerse V5.6 is a proprietary video generation large model developed by A-Soul Technology, featuring comprehensive upgrades in text-to-video and image-to-video capabilities. The model significantly improves image clarity, complex motion stability, and audio-visual synchronization, achieving more accurate lip-syncing and natural emotional expression in multi-character dialogue scenarios. It also optimizes composition, lighting, and texture consistency for further enhanced generation quality. PixVerse V5.6 ranks in the global top tier of the Artificial Analysis text-to-video and image-to-video leaderboards.
pixverse/pixverse-v5.6-kf2v
Achieves seamless transitions between any two images for smooth, visually striking scene shifts. PixVerse V5.6, developed by Artifactory, upgrades text-to-video and image-to-video capabilities. The model significantly improves visual clarity, complex motion stability, and audio-visual coordination, with accurate lip-sync and natural emotion in multi-character dialogues. It optimizes composition, lighting, and texture consistency for enhanced overall quality. PixVerse V5.6 ranks in the global top tier of Artificial Analysis text-to-video and image-to-video benchmarks.
pixverse/pixverse-v5.6-it2v
Upload any image and generate dynamic, coherent videos with customizable plots, pacing, and styles. PixVerse V5.6, developed by Artifactory, upgrades text-to-video and image-to-video capabilities. The model significantly improves visual clarity, complex motion stability, and audio-visual coordination, with accurate lip-sync and natural emotion in multi-character dialogues. It optimizes composition, lighting, and texture consistency for enhanced overall quality. PixVerse V5.6 ranks in the global top tier of Artificial Analysis text-to-video and image-to-video benchmarks.
MiniMax/speech-2.8-turbo
MiniMax speech large model intelligently predicts text emotions and intonation based on context to generate ultra-natural, high-fidelity, and personalized speech. Demonstrates strong capabilities in social interaction, podcasting, audiobooks, news, education, digital humans, and other scenarios.
MiniMax/speech-2.8-hd
MiniMax speech large model intelligently predicts text emotions and intonation based on context to generate ultra-natural, high-fidelity, and personalized speech. Demonstrates strong capabilities in social interaction, podcasting, audiobooks, news, education, digital humans, and other scenarios.
MiniMax/speech-02-turbo
MiniMax speech large model intelligently predicts text emotions and intonation based on context to generate ultra-natural, high-fidelity, and personalized speech. Demonstrates strong capabilities in social interaction, podcasting, audiobooks, news, education, digital humans, and other scenarios.
MiniMax/speech-02-hd
MiniMax speech large model intelligently predicts text emotions and intonation based on context to generate ultra-natural, high-fidelity, and personalized speech. Demonstrates strong capabilities in social interaction, podcasting, audiobooks, news, education, digital humans, and other scenarios.
gui-plus-2026-02-26
GUI-series foundation model for cross-platform (mobile/desktop) interface understanding and interaction, supporting complex multi-step tasks and multi-agent collaboration.
fun-asr-realtime-2026-02-28
This is the real-time version of Tongyi Lab's next-generation end-to-end speech recognition model, based on leading proprietary speech technology, and boasts exceptional contextual awareness and high-precision speech transcription capabilities. Based on an end-to-end architecture, Fun-ASR integrates innovative RAG technology, supporting multi-dimensional features such as large-scale hotword customization, automatic filtering of sensitive and modal particles, ITN normalization, and punctuation prediction, significantly improving overall recognition accuracy and contextual relevance. Furthermore, Fun-ASR supports flexible switching between Chinese and English, covers multiple regional dialects, and boasts enhanced noise robustness, adapting to diverse and complex environments.
qwen3-asr-flash-2026-02-10
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in multiple languages, while ensuring precise transcription even in complex audio environments.This version is a snapshot dated February 10, 2026.
qwen-image-2.0-pro-2026-03-03
The full-featured Qwen-Image-2.0 series models integrate image generation and image editing, offering enhanced text rendering with support for 1,000-token prompts, more refined realistic textures, detailed depiction of photorealistic scenes, and stronger semantic adherence. The full-featured version delivers text rendering and realistic textures in the 2.0 series.This version is a snapshot as of March 3, 2026.
qwen-image-2.0
qwen-image-2.0-2026-03-03
The Qwen-Image-2.0 series of accelerated models integrates image generation and image editing, offering enhanced text-rendering capabilities with support for 1,000-token prompts, more realistic textures, finely detailed photorealistic scenes, and improved semantic consistency. The accelerated version effectively strikes an optimal balance between model performance and quality.
qwen-flash-character-2026-02-26
The Qwen Role-Playing Model Series is specifically optimized for muti-language anthropomorphic interaction scenarios. It demonstrates advanced capabilities in character consistency maintenance, context-aware dialogue progression, and empathetic engagement, enabling precise personalized character embodiment. This version significantly enhances Japanese linguistic localization (including dialects and honorifics), human-like role-playing authenticity, narrative coherence control, and scenario-based cognitive intelligence.
cosyvoice-v3.5-plus
A high-expressiveness speech synthesis model in the CosyVoice series with enhanced voice cloning and design capabilities. Supports free-style instruction control while maintaining speaker similarity, offering rich synthesis styles. Reduces first-syllable latency, improves pronunciation accuracy, and enhances prosody and sound quality. Enables ultra-natural multilingual (Chinese, English, German, French, Russian, Japanese, Korean, Portuguese, Thai, Indonesian, Vietnamese) real-time speech synthesis.
cosyvoice-v3.5-flash
A high-performance speech synthesis model in the CosyVoice series with enhanced voice cloning and design capabilities. Supports free-style instruction control while maintaining speaker similarity, offering rich synthesis styles. Reduces first-syllable latency, improves pronunciation accuracy, and enhances prosody and sound quality. Enables ultra-natural multilingual (Chinese, English, German, French, Russian, Japanese, Korean, Portuguese, Thai, Indonesian, Vietnamese) real-time speech synthesis.
MiniMax-M2.5
MiniMax-M2.5 is MiniMax's open-source large model, trained on hundreds of thousands of real-world complex scenarios through large-scale reinforcement learning, achieving or surpassing industry SOTA in productivity scenarios like programming, tool calls, search, and office tasks.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3.5-35b-a3b
The Qwen3.5 Series 35B-A3B is a native vision-language model designed with a hybrid architecture that integrates linear attention mechanisms and a sparse mixture-of-experts model, achieving higher inference efficiency. Its overall performance is comparable to that of the Qwen3.5-27B.
qwen3.5-27b
The Qwen3.5 27B native vision-language Dense model incorporates a linear attention mechanism, delivering fast response times while balancing inference speed and performance. Its overall capabilities are comparable to those of the Qwen3.5-122B-A10B.
qwen3.5-122b-a10b
The Qwen3.5 122B-A10B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. In terms of overall performance, this model is second only to Qwen3.5-397B-A17B. Its text capabilities significantly outperform those of Qwen3-235B-2507, and its visual capabilities surpass those of Qwen3-VL-235B.
qwen3-coder-next
The new-generation code generation model in the Qwen3 series delivers performance close to that of Qwen3-Coder-Plus while offering even better capabilities. The model has been optimized with a focus on repository-level understanding, supports multi-turn tool interactions, and enhances its compatibility with agentic coding tools.
glm-5
GLM-5 targets coding and agent scenarios, achieving open-source SOTA in complex systems engineering with capabilities approaching Claude Opus, built on a 744B foundation with asynchronous reinforcement learning and sparse attention.
qwen3.5-plus
qwen3.5-plus-2026-02-15
The Qwen3.5 native vision-language series Plus models are built on a hybrid architecture that integrates linear attention mechanisms with sparse mixture-of-experts models, achieving higher inference efficiency. In a variety of task evaluations, the 3.5 series consistently demonstrates performance on par with state-of-the-art leading models. Compared to the 3 series, these models show a leap forward in both pure-text and multimodal capabilities.
qwen3.5-397b-a17b
The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. It delivers state-of-the-art performance comparable to leading-edge models across a wide range of tasks, including language understanding, logical reasoning, code generation, agent-based tasks, image understanding, video understanding, and graphical user interface (GUI) interactions. With its robust code-generation and agent capabilities, the model exhibits strong generalization across diverse agent.
qwen3-asr-flash-realtime-2026-02-10
The real-time version of Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in multiple languages, while ensuring precise transcription even in complex audio environments.This version is a snapshot dated February 10, 2026.
kimi/kimi-k2.5
Kimi K2.5 is Kimi's AI model released in 2026, achieving open-source SoTA performance in agentic tasks, coding, visual understanding, and a wide range of general intelligence tasks. It is also a model from Kimi, featuring a native multimodal architecture that supports visual and text inputs, thinking and non-thinking modes, and conversational and agentic tasks.
MiniMax/MiniMax-M2.5
MiniMax-M2.5 is MiniMax's open-source large model, trained on hundreds of thousands of real-world complex scenarios through large-scale reinforcement learning, achieving or surpassing industry SOTA in productivity scenarios like programming, tool calls, search, and office tasks.
MiniMax/MiniMax-M2.1
M2.1 was designed to break the barrier that top-tier agent capabilities are exclusive to closed-source models. Through targeted model-level optimizations, we have significantly enhanced its performance in code generation, tool calling, complex instruction following, and long-horizon planning tasks. MiniMax-M2.1 demonstrates exceptional stability across diverse scenarios, from automating multilingual software development to executing complex, multi-step office workflows. We are committed to providing developers with a fully transparent, controllable, and highly available foundation model for building next-generation autonomous agent applications.
fun-asr-flash-8k-realtime
fun-asr-flash-8k-realtime-2026-01-28
A lightweight real-time Mandarin speech recognition model optimized for Chinese call-center scenarios, supporting multi-dialect accents and achieving low-latency, high-accuracy transcription in low-sample-rate/low-SNR environments.
qwen3-tts-vd-2026-01-26
Qwen3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 26, 2026.
qwen3-tts-vc-2026-01-22
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated from the qwen-voice-enrollment service, and supports speech output in 11 languages using the same voice timbre. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 22, 2026.
qwen3-tts-instruct-flash
qwen3-tts-instruct-flash-2026-01-26
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. The Instruct model processes the synthesis effect through natural language, ensuring highly appropriate emotional and expressive speech in different contexts. Currently, it supports 25 timbres for both Chinese and English Instruct adjustments.
kimi-k2.5
Kimi-k2.5 is a model from Moonshot AI with native multimodal architecture, supporting visual/text inputs, thinking/non-thinking modes, and both dialogue and agent tasks.
wan2.6-r2v-flash
Wan2.6 reference to video flash, faster and more cost-effective generation. Supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances.
qwen3-vl-rerank
Qwen3-VL-Rerank is a reranking model that deeply understands rich multimodal information (text, images, video). After initial retrieval, it applies advanced cross-modal correlation to intelligently re-rank candidates, prioritizing the most relevant results. It improves cross-modal search accuracy, optimizes image/video retrieval precision, enhances image clustering quality, and enables efficient complex multimodal retrieval and tagging.
siliconflow/deepseek-v3.2
DeepSeek-V3.2 is a model combining high computational efficiency with outstanding reasoning and agent performance, built on three key breakthroughs: DeepSeek Sparse Attention (DSA), an efficient attention mechanism that significantly reduces computational complexity while preserving model performance and is specially optimized for long-context scenarios; a scalable reinforcement learning framework through which the model rivals GPT-5, with its high-compute version matching Gemini-3.0-Pro in reasoning; and a large-scale agent task synthesis pipeline that integrates reasoning into tool-use scenarios to improve instruction-following and generalization in complex interactive environments. The model achieved gold-medal performance at the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI).
siliconflow/deepseek-v3.1-terminus
DeepSeek-V3.1-Terminus is an updated version of DeepSeek-V3.1, maintaining the model's core capabilities while addressing user feedback through fixes and optimizations. It retains the same architecture as DeepSeek-V3 and achieves significant enhancements in specific domains.
siliconflow/deepseek-v3-0324
The new DeepSeek-V3 (DeepSeek-V3-0324) shares the same base model as the previous DeepSeek-V3-1226, with improvements focused solely on post-training methods. By incorporating reinforcement learning techniques used in training DeepSeek-R1, the new V3 model significantly enhances performance on reasoning tasks, outscoring GPT-4.5 on math and coding benchmarks. Additionally, the model demonstrates noticeable improvements in tool calling, role-playing, and conversational interactions.
siliconflow/deepseek-r1-0528
A large language model enhanced through extensive reinforcement learning during post-training, achieving strong reasoning capabilities with minimal labeled data. Performs well in mathematics, coding, and natural language reasoning tasks.
qwen3-vl-plus
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen3-max-2026-01-23
Compared with the snapshot as of September 23, 2025, the Qwen-3 series Max model in this release achieves an effective integration of thinking and non-thinking modes, resulting in a comprehensive and substantial improvement in the model's overall performance. In thinking mode, the model simultaneously supports web search, web information extraction, and a code interpreter tool, enabling it to tackle more complex and challenging problems with greater accuracy by leveraging external tools while engaging in slow, deliberative reasoning. This version is based on a snapshot taken on January 23, 2026.
MiniMax-M2.1
MiniMax-M2.1 is MiniMax's open-source large model, focusing on real-world complex tasks with core strengths in multilingual programming and long-chain agent capabilities.
qwen3-vl-flash-2026-01-22
The Qwen3 series of small-sized visual understanding models effectively integrates thinking and non-thinking modes. Compared with the snapshot taken on October 15, 2025, the overall performance of the model has improved significantly: it delivers enhanced capabilities in general visual recognition and reasoning, and shows marked improvements in recognition accuracy across various business scenarios such as security, in-store inspections, equipment monitoring, and photo-based problem solving. This version is a snapshot as of January 22, 2026.
qwen3-vl-embedding
Qwen3-VL-Embedding is a unified multimodal vector model based on Qwen3-VL, supporting text, image, video (single or mixed modality) input. It outputs unified representation vectors, suitable for cross-modal retrieval, image/video search, clustering, complex multimodal retrieval, and tagging.
qwen3-tts-instruct-flash-realtime
qwen3-tts-instruct-flash-realtime-2026-01-22
qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. The Instruct model processes the synthesis effect through natural language, ensuring highly appropriate emotional and expressive speech in different contexts. Currently, it supports 25 timbres for both Chinese and English Instruct adjustments. This model is equivalent to the snapshot version released on January 22, 2026.
wan2.6-i2v-flash
Wan2.6 image to video flash, faster and more cost-effective generation. Intelligent shot scheduling enables multi‑camera storytelling, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
qwen-image-edit-max
qwen-image-edit-max-2026-01-16
The Max series Qwen's image editing models delivers more stable and versatile editing capabilities: enhanced industrial design and geometric reasoning, improved character consistency, reduced offset issues, and integrated LoRA capabilities for a wider range of image editing functions.
qwen3-tts-vd-realtime-2026-01-15
Qwen3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 15, 2026.
qwen3-tts-vc-realtime-2026-01-15
Qwen 3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated from the qwen-voice-enrollment service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 15, 2026.
qwen-flash-character
The Qwen Role-Playing Model Series is specifically optimized for muti-language anthropomorphic interaction scenarios. It demonstrates advanced capabilities in character consistency maintenance, context-aware dialogue progression, and empathetic engagement, enabling precise personalized character embodiment. This version significantly enhances Japanese linguistic localization (including dialects and honorifics), human-like role-playing authenticity, narrative coherence control, and scenario-based cognitive intelligence.
tongyi-xiaomi-analysis-pro
Dialogue Analysis-Pro model for advanced complex analysis of intricate quality inspection rules with complex business logic, supporting fine-grained analysis standards and featuring enhanced multi-turn context modeling, deep semantic understanding, and reasoning.
tongyi-xiaomi-analysis-flash
Dialogue Analysis-Flash model for daily tasks like dialogue information extraction and scenario classification, supporting custom analysis standards with enhanced semantic understanding for low-latency offline/online analysis.
qwen-image-plus-2026-01-09
The Qwen series of image-generation models boasts exceptional text-rendering capabilities and excels in complex text rendering as well as a wide range of generation and editing tasks. This version, a snapshot taken on January 9, 2026, is a distilled and accelerated variant of Qwen-Image-Max, enabling faster generation of high-quality images.
2025
qwen-image-max
qwen-image-max-2025-12-30
The Max series of qwen's image generation model excels across a wide range of generation tasks. Compared with the Plus series, it significantly reduces the "AI-like" feel in generated images, enhancing their realism. It delivers more lifelike material textures for human subjects, finer and more detailed natural textures, and more visually appealing text rendering.
glm-4.7
Zhipu's model with enhanced coding and multi-step reasoning capabilities, supporting long-term task planning, tool collaboration, and immersive writing/role-playing.
z-image-turbo
Z-Image-Turbo is a highly efficient image-generation model that has topped the Artificial Analysis benchmark as the world's No. 1 open-source text-to-image model. With just 6 billion parameters and an 8-step inference process, it generates photo-realistic images comparable to those produced by large-scale commercial models, while excelling in bilingual Chinese–English text rendering, complex semantic understanding, and diverse thematic generation.
qwen3-vl-plus-2025-12-19
The Qwen3 series of visual understanding models effectively integrates thinking and non-thinking modes. Compared to the snapshot released on September 23, this version delivers superior performance in reasoning and analysis tasks as well as style control, while also offering lower latency and faster response speeds. This version is based on a snapshot taken on December 19, 2025.
wan2.6-r2v
Wan2.6 reference to video, supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances. Note: When a video is used as a reference, the input video is also billed. For details, see the model billing documentation.
wan2.6-t2i
Wan2.6 text to image, Upgraded visual quality, aesthetics, and instruction-following deliver precise style control, realistic portraits, long-text understanding, and broad historical/cultural IP coverage, enabling high-quality, highly expressive visual generation.
wan2.6-image
Wan2.6 Image, An all-round image generation model that supports joint text–image reasoning, multi-image creative fusion, commercial-grade consistency, aesthetic style transfer, and precise control of framing and lighting, significantly enhancing consistency, controllability, and expressiveness in image generation.
qwen-image-edit-plus-2025-12-15
The Qianwen series of Image Editing Plus models features enhanced character consistency, industrial design capabilities, and geometric reasoning abilities compared to the snapshot as of October 30. Additionally, it integrates LoRA capabilities such as lighting effects and effectively mitigates offset issues. This version is based on a snapshot taken on December 15, 2025.
qwen3-tts-vd-realtime-2025-12-16
Qwen 3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from December 16, 2025.
qwen-voice-design
Qwen Voice-Design model is a series of voice design models from Qwen Speech Model. It only requires a simple text description to quickly design a suitable voice. When used in conjunction with the qwen3-tts-vd-realtime model, it can design and output speech in 10 languages. Furthermore, the synthesized audio can adaptively adjust its tone based on the text and has good processing capabilities for complex text synthesis.
qwen3-omni-flash-realtime
qwen3-omni-flash-realtime-2025-12-01
The real-time version of the Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-omni-flash
qwen3-omni-flash-2025-12-01
Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-livetranslate-flash
qwen3-livetranslate-flash-2025-12-01
Qwen3-LiveTranslate-Flash is a high-precision, highly responsive, and robust multilingual real-time audio and video interpretation model. Leveraging Qwen3-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3-LiveTranslate-Flash provides both offline and real-time audio and video translation capabilities. It can understand 19 languages and speak 10 languages, and also supports 8 Chinese dialects.
wan2.6-t2v
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
wan2.6-i2v
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
deepseek-v3.2
DeepSeek-V3.2 is the official release of a model that incorporates DeepSeek Sparse Attention—a sparse attention mechanism. It's also the first model launched by DeepSeek that integrates reasoning into tool usage, supporting both reasoning-enabled and non-reasoning tool calls.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen3-tts-vc-realtime-2025-11-27
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated by the qwen3-voice-enrollment service, and supports speech output in 11 languages with the same voice timbre. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust the tone according to the text, and it also has good processing capabilities for complex text synthesis.This model is provided as a snapshot version.
qwen3-tts-flash-realtime
qwen3-tts-flash-realtime-2025-11-27
The Qwen3-TTS-Flash-Realtime model is Tongyi's real-time speech synthesis foundation model, featuring 17 expressive voices while delivering low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on extensive datasets, the system autonomously adjusts vocal tones based on text semantics and demonstrates robust capabilities for complex content synthesis.
qwen3-tts-flash
qwen3-tts-flash-2025-11-27
The Qwen3-TTS-Flash is Tongyi's offline text-to-speech foundation model, featuring 17 expressive voices while enabling low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on massive datasets, the system automatically adjusts vocal tones based on text semantics and demonstrates robust capabilities for synthesizing complex content.
qwen-voice-enrollment
The Qwen Voice-Enrollment model is a series of voice replication models from the qwen speech model. It can quickly replicate highly similar voices using audio of only 5 seconds or more. When used in conjunction with the qwen3-tts-vc-realtime model, it can replicate a person's voice with high fidelity and output speech in 10 languages. Furthermore, the synthesized audio can adaptively adjust its tone according to the text and has good processing capabilities for complex text synthesis.
fun-asr
The Fun ASR version, updated in April 2026, fully supports the seven major dialect systems of traditional Chinese (Mandarin/Wu/Xiang/Gan/Hakka/Min/Yue) and is compatible with over 20 regional Mandarin accents. It features specific optimizations for the rhythm, meter, and literary expression characteristics of classical Chinese poetry, improving the accuracy of poetry recognition and making it suitable for cultural heritage, educational explanations, and audiobooks. Improved punctuation prediction and text normalization capabilities make the output text more consistent with written expression habits; numbers, dates, and amounts are automatically converted to standard formats, enhancing readability and professionalism. The language support has been expanded to include English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, and Slovak, totaling 30 languages. This version is equivalent to the snapshot released on November 7, 2025.
qwen-vl-ocr
qwen-vl-ocr-2025-11-20
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen-mt-lite
Qwen-MT-Lite is a large language model of the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 32 languages at a cost-effective price. It offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen3-asr-flash-filetrans
qwen3-asr-flash-filetrans-2025-11-17
The large file transcription version of Qwen3-ASR-Flash. Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in multiple languages, ensuring precise transcription even in complex audio environments.
fun-asr-realtime-2025-11-07
This is the real-time version of Tongyi Lab's next-generation end-to-end speech recognition model, based on leading proprietary speech technology, and boasts exceptional contextual awareness and high-precision speech transcription capabilities. Based on an end-to-end architecture, Fun-ASR integrates innovative RAG technology, supporting multi-dimensional features such as large-scale hotword customization, automatic filtering of sensitive and modal particles, ITN normalization, and punctuation prediction, significantly improving overall recognition accuracy and contextual relevance. Furthermore, Fun-ASR supports flexible switching between Chinese and English, covers multiple regional dialects, and boasts enhanced noise robustness, adapting to diverse and complex environments.This is a snapshot released on November 7, 2025.
fun-asr-2025-11-07
The Fun ASR version, updated in April 2026, fully supports the seven major dialect systems of traditional Chinese (Mandarin/Wu/Xiang/Gan/Hakka/Min/Yue) and is compatible with over 20 regional Mandarin accents. It features specific optimizations for the rhythm, meter, and literary expression characteristics of classical Chinese poetry, improving the accuracy of poetry recognition and making it suitable for cultural heritage, educational explanations, and audiobooks. Improved punctuation prediction and text normalization capabilities make the output text more consistent with written expression habits; numbers, dates, and amounts are automatically converted to standard formats, enhancing readability and professionalism. The language support has been expanded to include English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, and Slovak, totaling 30 languages. This is a snapshot released on November 7, 2025.
cosyvoice-v3-flash
Synthesis Capabilities: CosyVoice-v3-Flash is a high-performance speech synthesis model in the CosyVoice series from Tongyi Labs, offering improved naturalness, timbre, prosody, and emotional expressiveness compared to previous versions. This model supports real-time streaming text-to-speech synthesis. Cloning Capabilities: CosyVoice-v3-Flash is also a speech cloning model in the CosyVoice series from Tongyi Labs. Compared to previous versions, it improves pronunciation accuracy and timbre similarity, and adds support for more less commonly spoken languages (German, Spanish, French, Italian, Russian, Japanese). It can quickly generate highly similar and naturally sounding custom voices from just 5-20 seconds of reference audio.
gui-plus
GUI-series foundation model for cross-platform (mobile/desktop) interface understanding and interaction, supporting complex multi-step tasks and multi-agent collaboration.
kimi-k2-thinking
Kimi-k2-thinking is a general agentic reasoning model developed by Moonshot AI, specialized in deep reasoning and multi-step tool integration to solve complex problems.
qwen-mt-flash
Qwen-MT-Flash, a large language model from the Qwen series, has been fully upgraded with the Qwen 3 architecture for significantly enhanced performance and translation quality. It provides rapid, cost-effective translation across 92 languages, while supporting advanced features such as terminology intervention, format preservation, and domain-specific adaptation. It is the ideal choice for applications requiring a powerful balance of speed, quality, and cost.
qwen-image-edit-plus
qwen-image-edit-plus-2025-10-30
The qwen series of image editing Plus models further optimizes inference performance and system stability based on the initial Edit model, significantly reducing the response time for image generation and editing. It also supports returning multiple images in a single request, greatly enhancing user experience.
qwen3-asr-flash-realtime
qwen3-asr-flash-realtime-2025-10-27
The real-time version of Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in 11 languages, while ensuring precise transcription even in complex audio environments.
qwen3-vl-32b-thinking
The largest dense model in the Qwen3-VL series, its reasoning version boasts multimodal reasoning capabilities second only to Qwen3-VL-235B-Thinking. It excels in STEM and math problem-solving, general image and video understanding, and achieves state-of-the-art performance in multimodal agent capabilities, making it ideal for complex multimodal reasoning tasks.
qwen3-vl-32b-instruct
The largest dense model in the Qwen3-VL series, in its non-inference version, delivers overall performance second only to Qwen3-VL-235B-Instruct. It excels in document recognition and comprehension, demonstrates strong spatial awareness and object identification capabilities, and achieves state-of-the-art performance in 2D visual detection and spatial reasoning. It is well-suited for complex perception tasks across a wide range of general-purpose scenarios.
qwen3-rerank
A text-ranking model trained on the Qwen LLM foundation performs relevance ranking for input queries and candidate documents. It supports over 100 languages and long-text inputs, and is suitable for applications such as text retrieval and RAG. Its performance is aligned with the open-source Qwen3-Rerank series models.
qwen2.5-vl-embedding
Qwen2-5-VL-Embedding is a unified multimodal vector model based on Qwen2.5-VL, supporting text, image, video (single or mixed modality) input. It outputs unified representation vectors, suitable for cross-modal retrieval, image/video search, clustering, complex multimodal retrieval, and tagging.
glm-4.6
GLM's new model with comprehensive capability improvements over 4.5, featuring a 200K context window.
qwen3-vl-flash
qwen3-vl-flash-2025-10-15
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-8b-thinking
The Thinking version of the 8B Dense model in the Qwen3-VL series consumes less GPU memory and is capable of performing multimodal understanding and reasoning. It supports extremely long contexts such as lengthy videos and documents, 2D/3D visual localization, and features comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities.
qwen3-vl-8b-instruct
The Instruct version of the 8B Dense model in the Qwen3-VL series requires less GPU memory and provides comprehensively upgraded image/video understanding, support for extremely long contexts like lengthy videos and documents, spatial awareness, and object recognition capabilities, making it suitable for tackling complex real-world tasks.
qwen3-vl-30b-a3b-thinking
The Thinking version of the second-largest MoE model in the Qwen3-VL series features fast response speeds and enhanced multimodal understanding and reasoning capabilities, visual agents, and support for extremely long contexts such as lengthy videos and documents. It also boasts comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities, making it well-suited for complex real-world tasks.
qwen3-vl-30b-a3b-instruct
The Instruct version of the second-largest MoE model in the Qwen3-VL series offers rapid response speeds and supports extremely long contexts like lengthy videos and documents. It includes comprehensively upgraded image/video understanding, spatial awareness, and object recognition capabilities, as well as 2D/3D visual localization, enabling it to handle intricate real-world challenges.
deepseek-v3.2-exp
Experimental version introducing DeepSeek Sparse Attention (a sparse attention mechanism), exploring optimization and validation for training and inference efficiency on long texts.
fun-asr-mtl
fun-asr-mtl-2025-08-25
Fun's multilingual speech recognition model supports over 31 languages and allows for free language switching, making it the top choice for users expanding overseas, especially to Southeast Asia. Fun-asr is an upgraded version of this model; switching to Fun-asr is recommended.
wan2.5-i2i-preview
The upgraded Wan2.5 Preview image edit model, newly upgraded model architecture supports rich image editing capabilities via instruction control, with enhanced instruction adherence. It also enables multi-image reference generation with high consistency and demonstrates excellent text generation performance.
tongyi-embedding-vision-plus
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
tongyi-embedding-vision-flash
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
qwen3-vl-plus-2025-09-23
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.This version is a snapshot as of September 23, 2025
qwen3-vl-235b-a22b-thinking
Qwen3 series VL models feature significantly enhanced multimodal reasoning capabilities, with a particular focus on optimizing the model for STEM and mathematical reasoning. Visual perception and recognition abilities have been comprehensively improved, and OCR capabilities have undergone a major upgrade.
qwen3-vl-235b-a22b-instruct
The Qwen3 series VL models has been comprehensively upgraded in areas such as visual coding and spatial perception. Its visual perception and recognition capabilities have significantly improved, supporting the understanding of ultra-long videos, and its OCR functionality has undergone a major enhancement.
qwen3-max
qwen3-max-2025-09-23
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-livetranslate-flash-realtime
qwen3-livetranslate-flash-realtime-2025-09-22
The real-time version of Qwen3-LiveTranslate-Flash, which is a high-precision, highly responsive, and robust multilingual simultaneous audio and video interpretation model. Leveraging Qwen3-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3-LiveTranslate-Flash offers both offline and real-time audio and video translation capabilities. It can understand 19 languages and speak 10 languages, including 8 Chinese dialects.
qwen3-coder-plus-2025-09-23
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability. This is a snapshot from 23 September, 2025.Compared to the previous version (snapshot from July 22), it demonstrates improved robustness in downstream task performance and tool invocation, along with enhanced code security.
qwen-vl-ocr-latest
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen-image-plus
The first image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
fun-asr-realtime
fun-asr-realtime-2025-09-15
This is the real-time version of Tongyi Lab's next-generation end-to-end speech recognition model, based on leading proprietary speech technology, and boasts exceptional contextual awareness and high-precision speech transcription capabilities. Based on an end-to-end architecture, Fun-ASR integrates innovative RAG technology, supporting multi-dimensional features such as large-scale hotword customization, automatic filtering of sensitive and modal particles, ITN normalization, and punctuation prediction, significantly improving overall recognition accuracy and contextual relevance. Furthermore, Fun-ASR supports flexible switching between Chinese and English, covers multiple regional dialects, and boasts enhanced noise robustness, adapting to diverse and complex environments.
qwen-image-edit
The first Qwen image editing model extends Qwen-Image's text rendering to editing tasks. It offers precise bilingual (Chinese/English) text editing, dual visual and semantic editing, and strong cross-benchmark performance.
wan2.5-t2v-preview
The upgraded Wan2.5 Preview text to video model, newly upgraded model architecture supports synchronized audio generation with visuals, enables 10-second long video generation, and offers enhanced instruction adherence, improved motion capabilities, and superior image quality.
wan2.5-t2i-preview
The upgraded Wan2.5 Preview text to image model, newly upgraded model architecture significantly enhances visual aesthetics, design sensibility, and realistic texture. It excels in precise instruction adherence, generates text proficiently in English, Chinese, and less common languages, and supports the generation of complex structured long texts, charts, and architectural diagrams.
wan2.5-i2v-preview
The upgraded Wan2.5 Preview image to video model, newly upgraded model architecture supports synchronized audio generation with visuals, enables 10-second long video generation, and offers enhanced instruction adherence, improved motion capabilities, and superior image quality.
wan2.2-animate-move
wan2.2-animate-move is a character animation generation model. Users simply upload a character photo and a reference performance video, and the model transfers the expressions and actions from the video onto the character in the image, producing a high-fidelity animated video.
wan2.2-animate-mix
wan2.2-animate-mix is a character replacement model product. By uploading a character photo and a performance video, users can accurately replace the character in the original video with the character from the photo, while completely preserving environmental details such as the scene, lighting, and color tone of the original video.
qwen3-omni-flash-realtime-2025-09-15
The real-time version of the Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-omni-flash-2025-09-15
Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.This version is a snapshot version from September 15, 2025.
qwen3-omni-30b-a3b-captioner
Qwen3-Omni-30b-a3b-Captioner is a powerful fine-grained audio analysis model designed to generate accurate and comprehensive content descriptions in complex and changing audio scenarios. It can automatically parse and describe various audio content, from complex speech and ambient sounds to music and film and television sound effects, and can maintain stable and reliable output even in multi-source and mixed environments.
qwen3-tts-flash-realtime-2025-09-18
The Qwen3-TTS-Flash-Realtime-2025-09-18 model is Tongyi's real-time speech synthesis foundation model, featuring 17 expressive voices while delivering low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on extensive datasets, the system autonomously adjusts vocal tones based on text semantics and demonstrates robust capabilities for complex content synthesis.This model is provided as a snapshot version.
qwen3-tts-flash-2025-09-18
The Qwen 3-TTS-Flash-2025-09-18 is Tongyi's offline text-to-speech foundation model, featuring 17 expressive voices while enabling low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on massive datasets, the system automatically adjusts vocal tones based on text semantics and demonstrates robust capabilities for synthesizing complex content. This model is provided as a snapshot version.
qwen-vl-ocr-2025-08-28
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
wan2.2-kf2v-flash
The all-new Wan2.2 First and Last Frame to video model is here. We've optimized motion stability and success rates, enhanced prompt adherence, and enabled seamless transitions between two images.
qwen3-next-80b-a3b-thinking
A new generation of Qwen3-based open-source thinking mode models. This version offers improved instruction following and streamlined summary responses over the previous iteration (Qwen3-235B-A22B-Thinking-2507).
qwen3-next-80b-a3b-instruct
A new generation of open-source, non-thinking mode model powered by Qwen3. This version demonstrates superior Chinese text understanding, augmented logical reasoning, and enhanced capabilities in text generation tasks over the previous iteration (Qwen3-235B-A22B-Instruct-2507).
qwen-plus-2025-09-11
As of the September 11, 2025 snapshot, this release features enhanced instruction following and streamlined summary responses in thinking mode. Non-thinking mode provides superior Chinese text understanding and augmented logical reasoning capabilities. The model supports a 1M context length with tiered pricing.
qwen3-asr-flash
qwen3-asr-flash-2025-09-08
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in 11 languages, ensuring precise transcription even in complex audio environments.
qwen3-max-preview
A preview version of the Max model in the Qwen 3 series, achieving an effective integration of thinking and non-thinking modes. In thinking mode, there is a significant enhancement in capabilities such as intelligent agent programming, common-sense reasoning, and reasoning across mathematics, science, and general domains.
cosyvoice-v3-plus
Cloning capability: CosyVoice-v3-plus is a large voice cloning model in the CosyVoice series from Tongyi Lab. It offers superior sound quality and cloning fidelity, ideal for professional scenarios. With just 5-20 seconds of reference audio, it can rapidly generate a highly similar and natural-sounding custom voice. Synthesis capability: CosyVoice-v3-plus is a large speech synthesis model in the CosyVoice series from Tongyi Lab. It features enhanced sound quality and expressiveness, ideal for professional scenarios. The model supports real-time, streaming text-to-speech synthesis.
wan2.2-s2v-detect
Auxiliary model for wan2.2-s2v to validate input portrait images against required specifications, ensuring compatibility with video generation via wan2.2-s2v.
wan2.2-s2v
High-quality Video Generation model that creates dynamic character videos (speaking/singing/performing) from input character images and voice audio files.
deepseek-v3.1
A hybrid inference architecture model supporting both thinking mode and non-thinking mode, featuring higher reasoning efficiency and stronger agent capabilities.
qwen-mt-image
Qwen-MT-Image specializes in image translation, converting images across 11 languages (Chinese, English, Japanese, etc.) to target languages. It accurately preserves layout and content, supporting custom features like terminology definitions, sensitive word filtering, and product detection for flexible, accurate image localization.
qwen-deep-research
Qwen-Deep-Research is an advanced agent system for complex research tasks, equipped with multi-step reasoning and global planning capabilities. It leverages tools like internet search to perform detailed task decomposition, reasoning, and analysis, ultimately generating traceable, logically rigorous research reports.
fun-asr-2025-08-25
This is a new-generation large speech recognition model that focuses on Mandarin Chinese, English, and Japanese. It supports a wide range of regional dialects, offers enhanced noise robustness, and adapts to diverse and complex environments.This is a snapshot released on August 25, 2025.
qwen-image
The first image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
wan2.2-i2v-flash
The upgraded Wan 2.2 image to video Flash model, delivers faster speed with optimized stability, more powerful prompt following, improved consistency for text, portraits, and products, and precise shot control.
glm-4.5-air
GLM-4.5-Air uses MoE architecture (106B total, 12B active parameters), a compact variant of GLM-4.5 for resource-constrained scenarios.
glm-4.5
GLM-4.5 adopts a Mixture-of-Experts (MoE) architecture with 355B total parameters and 32B active parameters, excelling in complex reasoning, code generation, and agent interactions.
qwen3-coder-flash
qwen3-coder-flash-2025-07-28
Based on Qwen3, this code generation model inherits the coding agent capabilities of Qwen3-Coder-Plus and supports multi-turn tool interaction. It features focused optimizations on repository-level understanding and enhanced tool-calling stability.
qwen-flash
qwen-flash-2025-07-28
The Qwen3 Flash model offers a powerful fusion of thinking and non-thinking modes with dynamic in-conversation switching, excelling in complex reasoning while showing significant gains in instruction following and text comprehension. It supports a 1M context length and is billed on a tiered model corresponding to context usage.
qwen3-coder-30b-a3b-instruct
Qwen3-based code generation model that inherits the coding agent ability of Qwen3-Coder-480B-A35B-Instruct; code capability reaches SOTA at the same scale.
qwen-plus-latest
The Qwen series model optimized for balanced performance, offering inference efficiency between Qwen-Max and Qwen-Turbo, is designed to handle moderately complex tasks effectively. This dynamically updated version implements changes without prior notice.
qwen3-30b-a3b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-30B-A3B) excels in complex thinking tasks, including logic, math, science, code, and other challenging scenarios; instruction following, text understanding, and multilingual translation capabilities significantly improved.
qwen-plus-2025-07-28
Qwen3 series Plus model, integrates thinking and non-thinking modes and can switch modes during dialogue. Compared to the prior version, adds dedicated enhancements for Chinese & English capabilities and tool calling. This is a snapshot from 28 July, 2025; first to support 1 M context length, and uses tiered pricing.
qwen3-30b-a3b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-30B-A3B) shows major improvements in Chinese, English, and overall multilingual general capabilities. Optimized for subjective open-ended tasks, delivering responses significantly more aligned with user preferences and more helpful.
wan2.2-t2v-plus
The upgraded Wan 2.2 Plus text to video model, delivers higher quality results with stable sweeping complex movements, cinematic vision control, more powerful prompt following, and realistic world recreation.
wan2.2-t2i-plus
The upgraded Wan 2.2 Plus text to image model, delivers richer image detail with enhanced creativity, stability, and realism. It also features stronger prompt following and native support for multiple styles. Up to 2 million pixel generation and prompt enhancement are supported as well.
wan2.2-t2i-flash
The upgraded Wan 2.2 Flash text to image model, delivers faster speed with enhanced creativity, stability, and realism. It also features stronger prompt following and native support for multiple styles. Up to 2 million pixel generation and prompt enhancement are supported as well.
wan2.2-i2v-plus
The upgraded Wan 2.2 Plus image to video model, delivers higher quality results with optimized stability, more powerful prompt following, improved consistency for text, portraits, and products, and precise shot control.
qwen3-235b-a22b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-235B-A22B) shows major improvements in logical ability, general capabilities, knowledge enhancement, and creativity, suitable for high-difficulty, strong-thinking scenarios.
qwen-doc-turbo
Qwen-Doc-Turbo rapidly extracts precise information from documents, supporting tagging, classification, content moderation, and summarization.
qwen3-coder-plus
qwen3-coder-plus-2025-07-22
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability.
qwen3-coder-480b-a35b-instruct
Qwen3-based code generation model with strong coding agent power; code capability reaches open-source SOTA.
qwen3-235b-a22b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-235B-A22B) shows slight improvements in subjective creativity and model safety.
qwen-mt-turbo
Qwen-MT-Turbo is a large language model within the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 92 languages at a cost-effective price point. It also offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen-mt-plus
Qwen-MT-Plus, the translation model from our Qwen series, is now fully upgraded with the Qwen3 architecture. It supports 92 languages and delivers exceptionally accurate and natural-sounding translations. Its advanced capabilities in contextual understanding, terminology control, and format preservation make it a superior choice over traditional models, especially for specialized domains.
qwen-tts-realtime
qwen-tts-realtime-latest
qwen-tts-realtime-2025-07-15
Qwen-TTS-Realtime is a speech synthesis model in the Qwen series with bidirectional context awareness. It enables low-latency, high-fidelity generation of multi-voice, dialectal, and long-text bidirectional streaming.
qwen-plus-2025-07-14
The Plus model of the Qwen3 Series, achieving effective integration of thinking mode and non-thinking mode, allows switching modes during conversations. This is a snapshot from July 14, 2025. Compared to the previous version, there has been a significant improvement in both Chinese and English capabilities under non-thinking mode, with enhanced tool-calling abilities.
Moonshot-Kimi-K2-Instruct
Kimi-K2 is Moonshot's first open-source trillion-parameter MoE model in China, featuring 32B activated parameters with exceptional coding and tool-calling capabilities.
qwen-tts-latest
qwen-tts-2025-05-22
Qwen-TTS is the first speech synthesis model in the Qwen series, supporting Chinese, English, and mixed inputs. It adaptively adjusts output tone based on text, offers natural voice quality, and supports streaming output.
qwen-turbo
The Turbo model of the Qwen3 series. It effectively integrates thinking mode and non-thinking mode, allowing for mode switching during conversations. Its reasoning capabilities rival those of QwQ-32B with a smaller parameter size, while its general capabilities significantly surpass those of Qwen2.5-Turbo, achieving the SOTA level in the same scale within the industry.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
deepseek-r1
A large language model enhanced through extensive reinforcement learning during post-training, achieving strong reasoning capabilities with minimal labeled data. Performs well in mathematics, coding, and natural language reasoning tasks.
qwen-vl-plus
Qwen-VL-Plus is the enhanced version of the large visual language model. It significantly improves detail recognition and text recognition capabilities, supporting images with resolutions exceeding one million pixels and any aspect ratio specifications. The model delivers exceptional performance across a wide range of visual tasks.
text-embedding-v4
The General Text Vector V4 version is a multi-language text vector model developed by the Tongyi Lab based on Qwen3. Compared to the V3 version, it significantly improves performance in text retrieval, clustering, and classification tasks. It achieves a 15% to 40% improvement in evaluation tasks such as MTEB multilingual, Chinese-English, and code retrieval. Additionally, it supports user-defined vector dimensions ranging from 64 to 2048.
deepseek-r1-0528
A large language model enhanced through extensive reinforcement learning during post-training, achieving strong reasoning capabilities with minimal labeled data. Performs well in mathematics, coding, and natural language reasoning tasks.
qvq-plus
Qwen QVQ Visual Reasoning Model Plus version supporting visual input and chain-of-thought output with enhanced capabilities in mathematics, programming, visual analysis, creation, and general tasks.
cosyvoice-v2
A generative speech synthesis model combining text understanding and speech generation through large-scale pretrained language models, supporting real-time streaming text-to-speech synthesis.
qwen-vl-max
Qwen-VL-Max is a large-scale visual language model of the Qwen series. Compared to the Plus version, it further enhances visual reasoning capabilities and instruction-following abilities, offering higher levels of visual perception and cognition. It delivers optimal performance on more complex tasks.
wanx2.1-vace-plus
Wanxiang 2.1 - Video Editing Unified Model - Plus. Supports localized editing, video redrawing, background expansion, duration extension, and reference-based generation. Enables multi-modal control through text, images, or videos.
qwen-omni-turbo-realtime
qwen-omni-turbo-realtime-latest
qwen-omni-turbo-realtime-2025-05-08
The real-time version of Qwen's new large multimodal understanding and generation model, suitable for real-time audio interaction scenarios. It supports the understanding of audio accompanied by text, images, and video mixed inputs, and can simultaneously generate speech and text in stream, providing four natural tones.
qwen3-8b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-7B.
qwen3-32b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-32B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-30b-a3b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability rivals QwQ-32B with a smaller parameter size, and its general capability significantly surpasses Qwen2.5-14B, reaching the SOTA level in the same scale industry.
qwen3-235b-a22b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-72B-Instruct, reaching the SOTA level in the same scale industry.
qwen-plus-2025-04-28
The Plus model of the Qwen3 series, effectively integrates thinking mode and non-thinking mode, allowing for mode switching during conversations. Its reasoning capabilities significantly surpass those of QwQ, and its general capabilities notably exceed those of Qwen2.5-Plus, reaching the SOTA level in the same scale within the industry. This model is the snapshot from April 28, 2025.
qwen3-14b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-14B.
aitryon-plus
A high-quality virtual try-on image generation model that produces try-on effects with enhanced image clarity, garment texture details, and logo reconstruction compared to aitryon. Requires longer generation time, suitable for non-time-critical scenarios.
qwen-vl-ocr-2025-04-13
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
wanx2.1-kf2v-plus
Wanxiang 2.1 - Keyframe-to-Video - Plus. Generates smooth transition videos from two input images. Supports complex large-scale motion, physics adherence, diverse artistic styles, and cinematic-grade visual quality. Enhanced instruction-following capability and richer visual details.
qwen-tts
qwen-tts-2025-04-10
Qwen-TTS is the first speech synthesis model in the Qwen series, supporting Chinese, English, and mixed inputs. It adaptively adjusts output tone based on text, offers natural voice quality, and supports streaming output.
qwen-omni-turbo-2025-03-26
The new multimodal understanding and generation model, supporting text, image, speech, and video input as well as mixed input understanding. It can simultaneously stream text and speech, with significantly enhanced speed for multimodal content understanding. Additionally, it offers four natural dialogue tones. This model is the snapshot from March 26, 2025, with significant improvements on visual capabilities over the snapshot of January 19, 2025.
qwen2.5-omni-7b
The new multi-modal understanding and generation model trained based on Qwen2.5. It supports text, image, speech, video, and mixed input understanding and can simultaneously generate streams of text and speech, significantly improves the speed of multi-modal content understanding. It provides four natural tones.
qvq-max
The Tongyi Qianwen QVQ visual reasoning model supports visual input and chain-of-thought output, demonstrating stronger capabilities in mathematics, programming, visual analysis, creation, and general tasks.
wanx2.1-imageedit
Image Editing model supporting preset and command-based tasks, including global/local editing (style transfer, inpainting, expansion, super-resolution) and reference-based generation.
sambert-zhiyuan-v1
Provides text-to-speech service combining SAMBERT+NSFGAN deep neural network algorithms with traditional domain expertise, featuring accurate pronunciation, natural prosody, high voice fidelity, and strong expressiveness.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
gte-rerank-v2
A multilingual text reranking model for semantic retrieval and RAG, ranking candidate documents by relevance to queries.
qwen-long-latest
qwen-long-2025-01-25
Qwen-Long is a large language model designed for ultra-long context processing, supporting Chinese, English, and other languages. It handles up to 10 million tokens (approximately 15 million characters or 15,000 document pages) in dialogues. Integrated with document services, it supports parsing and dialogue for text files (TXT, DOCX, PDF, XLSX, EPUB, MOBI, MD, CSV) and image files (BMP, PNG, JPG/JPEG, GIF, PDF scans). Notes: HTTP requests support up to 1M tokens; file submission is recommended for longer content.
sambert-eva-v1
sambert-beth-v1
sambert-waan-v1
sambert-betty-v1
sambert-brian-v1
sambert-cally-v1
sambert-cindy-v1
sambert-clara-v1
sambert-donna-v1
sambert-hanna-v1
sambert-indah-v1
sambert-perla-v1
sambert-zhida-v1
sambert-zhide-v1
sambert-zhimo-v1
sambert-zhina-v1
sambert-zhiqi-v1
sambert-zhiru-v1
sambert-zhiya-v1
sambert-zhiye-v1
sambert-camila-v1
sambert-zhichu-v1
sambert-zhifei-v1
sambert-zhigui-v1
sambert-zhihao-v1
sambert-zhijia-v1
sambert-zhilun-v1
sambert-zhimao-v1
sambert-zhinan-v1
sambert-zhishu-v1
sambert-zhiwei-v1
sambert-zhiyue-v1
sambert-zhijing-v1
sambert-zhiming-v1
sambert-zhiqian-v1
sambert-zhishuo-v1
sambert-zhiting-v1
sambert-zhixiao-v1
sambert-zhiying-v1
sambert-zhixiang-v1
sambert-zhistella-v1
sambert-zhimiao-emo-v1
Provides text-to-speech service combining SAMBERT+NSFGAN deep neural network algorithms with traditional domain expertise, featuring accurate pronunciation, natural prosody, high voice fidelity, and strong expressiveness.
cosyvoice-v1
A next-generation generative speech synthesis model integrating text understanding and speech generation via large-scale pretrained language models, supporting real-time streaming text-to-speech synthesis.
qwen-omni-turbo-2025-01-19
The new multimodal understanding and generation model, supporting text, image, speech, and video input as well as mixed input understanding. It can simultaneously stream text and speech, with significantly enhanced speed for multimodal content understanding. Additionally, it offers four natural dialogue tones.
qwq-plus
The enhanced version of the Qwen QwQ reasoning model, trained on the Qwen2.5 model, has significantly improved its reasoning capabilities through reinforcement learning. The model's core metrics in mathematics and coding (e.g., AIME 24/25, LiveCodeBench) as well as some general metrics (e.g., IFEval, LiveBench) have reached the level of the full version of DeepSeek-R1.
gummy-realtime-v1
A multilingual speech-to-text and translation model providing high-accuracy real-time transcription for 10 languages (including Chinese/English/Japanese/Korean) with cross-lingual translation.
gummy-chat-v1
A multilingual speech-to-text and translation model supporting real-time recognition of 10 mixed languages (60s limit) with Chinese/English/Japanese/Korean mutual translation.
wanx2.1-i2v-turbo
Image-to-Video Generation-Turbo model that transforms images into dynamic videos with complex movements and cinematic aesthetics, featuring faster generation speed and enhanced instruction compliance.
qwen-omni-turbo
qwen-omni-turbo-latest
The new multimodal understanding and generation model, supporting text, image, speech, and video input as well as mixed input understanding. It can simultaneously stream text and speech, with significantly enhanced speed for multimodal content understanding. Additionally, it offers four natural dialogue tones.
deepseek-r1-distill-qwen-7b
A distilled large language model based on Qwen2.5-Math-7B, trained using DeepSeek R1's outputs.
deepseek-r1-distill-qwen-32b
A distilled large language model based on Qwen2.5-32B, trained using DeepSeek R1's outputs.
deepseek-r1-distill-qwen-14b
A distilled large language model based on Qwen2.5-14B, trained using DeepSeek R1's outputs.
deepseek-r1-distill-qwen-1.5b
A distilled large language model based on Qwen2.5-Math-1.5B, trained using DeepSeek R1's outputs.
qwen-plus-2025-01-25
The Qwen series of models, which are well-balanced in capabilities, offer reasoning performance and speed that fall between Qwen-Max and Qwen-Turbo, making them suitable for moderately complex tasks. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
deepseek-v3
A self-developed Mixture-of-Experts (MoE) model with 671B parameters (activating 37B), pre-trained on 14.8T tokens. Demonstrates excellent capabilities in long-text processing, coding, mathematics, encyclopedic knowledge, and Chinese language tasks.
wanx2.1-i2v-plus
Image-to-Video Generation-Plus model that converts images into dynamic videos with complex motions, physical realism, cinematic quality, and improved instruction adherence for higher video fidelity.
wanx2.0-t2i-turbo
Enhanced Text-to-Image model specializing in realistic portraits and creative designs, upgraded in aesthetics, realism, and artistic quality with up to 2MP resolution and smart prompt rewriting support.
emoji-v1
Emoji is a facial animation video generation model that creates facial animation videos using face images and preset dynamic templates.
emoji-detect-v1
Emoji-Detect is an image detection model assisting emoji generation, used to detect whether character images meet video generation requirements.
qwen-plus-0112
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
aitryon-parsing-v1
An image segmentation model serving as a supporting module for AI try-on OutfitAnyone, capable of segmenting model images and garment images for pre/post-processing of try-on images.
wanx2.1-t2v-turbo
Wanxiang 2.1 - Text-to-Video - Turbo. High-speed video generation with support for complex motion, physics simulation, artistic styles, and cinematic quality. Improved instruction-following capability.
wanx2.1-t2v-plus
Wanxiang 2.1 - Text-to-Video - Plus. Generates high-quality videos from text prompts. Supports complex motion, realistic physics simulation, diverse artistic styles, and cinematic visuals. Enhanced instruction-following performance.
wanx2.1-t2i-turbo
Wanxiang 2.1 - Text-to-Image - Turbo. Accelerated generation speed with enhanced aesthetics, realism, and artistry. Maintains strong semantic understanding, style diversity, and 2MP resolution support. Includes smart prompt rewriting functionality.
wanx2.1-t2i-plus
Wanxiang 2.1 - Text-to-Image - Plus. Upgraded image generation with enhanced aesthetics, realism, and artistry. Improved semantic understanding, style generalization, and support for up to 2MP resolution. Features smart prompt rewriting and richer visual details.
2024
paraformer-realtime-8k-v2
Paraformer real-time speech recognition model supporting multilingual switching for live streaming/meetings. Enables language selection via language_hints parameter for enhanced accuracy in 8kHz customer service scenarios. Supports: Mandarin (including dialects), English, Japanese, Korean.
qwen-plus-1220
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
multimodal-embedding-v1
Multimodal vector model developed by Tongyi Lab based on pre-trained multimodal foundation models. Generates high-dimensional continuous vectors from text, image, or video inputs for downstream tasks including search, classification, and content moderation.
paraformer-8k-v2
Paraformer Mandarin speech recognition model with enhanced architecture, improved recognition accuracy, and 8kHz telephone speech support (Chinese hotword only).
tongyi-intent-detect-v3
Intent Detection and Slot Filling model for dialogue systems, enabling joint prediction of API-based intents and slot parameters in a single output, returning standardized JSON results with multiple API commands and filled slots.
videoretalk
Character Video Generation model that synthesizes lip-synced videos based on input character videos and voice audio, matching mouth movements to the audio content.
animate-anyone-template-gen2
A motion template generation model assisting AnimateAnyone, capable of extracting human motions from videos and creating templates.
animate-anyone-gen2
A video generation model that creates full-body motion videos based on human portraits and motion templates.
animate-anyone-detect-gen2
An image detection model assisting AnimateAnyone, used to verify if human figures in images meet video generation requirements.
qwen-vl-ocr-1028
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen-coder-plus
Qwen-Coder-Plus is a specialized language model for programming and code generation, delivering excellent performance and outstanding results.
liveportrait-detect
LivePortrait-detect is an auxiliary image detection model for assessing whether human figures in images meet video generation requirements.
liveportrait
LivePortrait is a video generation model that creates lightweight dynamic portrait videos from static images.
emo-v1
EMO is a video generation model that generates high-quality dynamic portrait videos based on character images.
emo-detect-v1
EMO-Detect is an image detection model assisting EMO, used to detect whether character images meet video generation requirements.
qwen-max
Qwen-Max supports a parameter scale of hundreds of billions and multiple input languages such as Chinese and English. Qwen-Max is updated in a rolling manner. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen-math-turbo
Qwen-Math-Turbo is a specialized language model for math problem-solving, featuring high inference speed and low cost.
qwen-math-plus-0919
qwen-math-plus-latest
Qwen-Math-Plus is a powerful math problem-solving model excelling in Chinese and English mathematical tasks, including equations, calculations, and proofs.
qwen-coder-turbo
Qwen-Coder-Turbo is a specialized language model for programming and code generation, featuring high inference speed and low cost.
voice-enrollment
A large-model voice replication service used in conjunction with Cosyvoice-v3. Utilizing advanced large-model technology for feature extraction, it can replicate voices without a training process. Only a very short audio clip is required to quickly generate a highly similar and natural-sounding custom voice.
video-style-transform
Video Style Repainting model that applies artistic styles (e.g., Japanese comic, 3D cartoon) to input video frames while preserving original appearances, generating stylized videos with diverse visual effects.
speech-biasing
Speech-Biasing allows users to define hotwords (specific terms or phrases) with prioritized recognition/translation. This improves accuracy for domain-specific terms in speech recognition and translation tasks.
qwen-math-plus
qwen-math-plus-0816
Qwen-Math-Plus is a powerful math problem-solving model excelling in Chinese and English mathematical tasks, including equations, calculations, and proofs.
cosyvoice-clone-v1
A voice cloning model leveraging advanced large model technology for feature extraction without training. Generates highly similar, natural-sounding custom voices from short audio samples.
paraformer-v2
Paraformer speech recognition model supporting multilingual recognition with language_hints parameter. Supports all sampling rates and hotwords across: Mandarin (including dialects), English, Japanese, Korean.
paraformer-realtime-v2
Paraformer real-time speech recognition model supporting multilingual switching with language_hints parameter. Supports all sampling rates and hotwords across: Mandarin (including dialects), English, Japanese, Korean.
image-instance-segmentation
Person instance segmentation employs detection and segmentation techniques to identify objects in images and generate pixel-level masks for precise object boundary delineation.
image-erase-completion
Image erasure and completion tool removing specified elements (people, objects, text, watermarks) while preserving backgrounds using computer vision and AIGC inpainting techniques.
text-embedding-v3
The general text vectorization model is a multilingual unified text vectorization model developed by Tongyi Lab based on the large language model (LLM) foundation. This model is designed for multiple mainstream languages worldwide and provides advanced vectorization services to help developers convert text data into high-quality vector data.
aitryon-refiner
An image refinement module for secondary generation of AI try-on results, producing higher-fidelity try-on images with improved realism.
wanx-virtualmodel
Virtual Model that replaces models and backgrounds in product images with virtual alternatives while maintaining poses, supporting interactive products (e.g., apparel, footwear, accessories).
virtualmodel-v2
Virtual Model that intelligently replaces models and backgrounds in product images while retaining poses, enabling diverse showcase of interactive products (e.g., apparel, accessories) with virtual alternatives.
wanx-poster-generation-v1
Creative Poster Generation model that automatically creates poster layouts and typography based on user input, supporting diverse styles (promotion, greeting) for personalized designs without design expertise.
shoemodel-v1
ShoeModel-V1 supports AI try-on for footwear by inputting multi-view shoe images and reshaping shoe regions in template images. It generates natural layouts, rich details, and realistic try-on results for applications like model product design, AI shoe try-on, and layout refinement.
wanx-sketch-to-image-lite
Doodle-to-Image model that generates artistic doodles from hand-drawn sketches and text descriptions, offering 5 styles (flat illustration, oil painting, etc.) for creative, educational, and design applications.
paraformer-realtime-v1
Paraformer Mandarin real-time speech recognition model for 16kHz+ live streaming/meeting scenarios.
paraformer-realtime-8k-v1
Paraformer Mandarin real-time speech recognition model optimized for 8kHz customer service scenarios.
wanx-x-painting
Image Local Repainting model based on the Composer framework, generating semantically consistent partial edits with diverse styles using input images, masks, and text prompts, delivering natural layouts with rich details.
image-out-painting
Image expansion model for free-form image extension with rotation support. Allows expansion via expansion ratio or pixel count parameters for creative design, auxiliary drawing, and film post-production applications.
aitryon
A high-performance virtual try-on image generation model that produces try-on effect images based on garment flat lay photos and human frontal full-body portraits. It generates try-on images quickly, suitable for time-sensitive scenarios.
qwen-long
Qwen-Long is a large language model designed for ultra-long context processing, supporting Chinese, English, and other languages. It handles up to 10 million tokens (approximately 15 million characters or 15,000 document pages) in dialogues. Integrated with document services, it supports parsing and dialogue for text files (TXT, DOCX, PDF, XLSX, EPUB, MOBI, MD, CSV) and image files (BMP, PNG, JPG/JPEG, GIF, PDF scans). Notes: HTTP requests support up to 1M tokens; file submission is recommended for longer content.
farui-plus
Tongyi FaRui is a legal industry large model trained on legal data and knowledge, integrating fine-tuning, reinforcement learning, RAG retrieval augmentation, and legal Agent technology, capable of answering legal questions, case analysis, legal document generation, and contract review.
wordart-texture
WordArt Jinshu - Text Texture Generation. Applies material and texture overlays to text content or images based on prompts, achieving 3D effects or scene integration. Produces high-quality artistic text suitable for poster backgrounds.
wordart-semantic
WordArt Jinshu - Text Deformation. Creatively morphs input text outlines based on prompts, enabling versatile typographic designs. Outputs black-background white-mask images with transformed text content.
text-embedding-v2
Universal Text Vector, a multilingual text embedding model developed by Tongyi Lab based on the LLM foundation, providing high-quality vector services for major global languages to convert text data into high-quality vector representations.
text-embedding-v1
Text-Embedding-V1 is a multilingual text vector model based on Qwen's LLM foundation, providing high-quality vector services for major global languages. It helps developers convert text data into high-quality vectors efficiently.
text-embedding-async-v2
Text-Embedding-Async-V2 is a batch processing interface for general text vectors. Users submit bulk vector computation requests via text, and the platform stores results in downloadable files after processing.
text-embedding-async-v1
Text-Embedding-Async-V1 is a batch processing interface for general text vectors. Users submit bulk vector computation requests via text, and the platform stores results in downloadable files after processing.
paraformer-v1
Paraformer bilingual Mandarin-English speech recognition model for 16kHz+ audio/video processing.
paraformer-mtl-v1
Paraformer multilingual speech recognition model supporting audio/video with 16kHz+ sampling rates across 14 languages/dialects including Mandarin, Cantonese, Wu, Minnan, and major global languages.
paraformer-8k-v1
Paraformer speech recognition file transcription API for converting common audio/video files to text, supporting 8kHz Mandarin telephone speech recognition.
facechain-generation
Generates portraits of trained character models in preset styles (e.g., ID photos, business portraits).
facechain-facedetect
Detects user-uploaded character images to determine if faces meet FaceChain fine-tuning standards, evaluating face count, size, angle, lighting, clarity, etc., supporting batch input with per-image results.
wanx-style-repaint-v1
Portrait Style Repainting model that transforms input portraits into artistic styles (e.g., watercolor, oil painting) while preserving original facial features.
wanx-background-generation-v2
Image Background Generation model that extends foreground images with natural lighting and realistic details, supporting text/image-guided generation and intelligent text overlay.
wanx-v1
Multilingual Text-to-Image Generation model supporting Chinese/English inputs, covering key styles including watercolor, oil painting, Chinese ink painting, sketch, flat illustration, anime, and 3D cartoon.
Singapore
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
kimi-k3.1-pro
Kimi-K3.1-Pro is a 3T-class open-weight model in Moonshot AI's K3.1 series, developed through large-scale post-training of Kimi K3. It supports a one-million-token context window, text and image inputs, text output, and native visual understanding. It is optimized for coding, 3D creation, financial research, and long-running agent tasks. It scores 65.7% on Terminal-Bench 4.0 and 64.4% on OfficeQA Pro v2, while the claim-level hallucination rate decreases from 23.4% for K3 to 8.3%.
qwen-audio-3.1-realtime-flash
A real-time duplex speech model with text and audio input and output, function calling, web search, and voice cloning. The context window is 262,144 tokens.
qwen-audio-3.1-tts-next
Qwen-Audio-3.1-TTS-Next is an AudioGen model designed for unified audio generation. Going beyond conventional TTS systems that focus solely on speech synthesis, it can generate complete audio content in a single pass from inputs such as text, timestamps, and reference audio, seamlessly combining speech, sound effects, ambient sounds, and more. The model supports a wide range of tasks, including multilingual TTS, single-speaker speech, multi-speaker dialogue, podcasts, cinematic soundscapes, ambient audio, and sound effects. Its key strengths include more natural and expressive speech, consistent voice identity, flexible timing control, and high-quality soundscape generation. These capabilities make it well suited for content creation, short-form video, podcasts, film and television production, game audio, and other professional audio creation scenarios.
qwen-audio-3.1-tts-plus
Qwen-Audio-3.1-TTS-Plus is a high-quality, professional text-to-speech model for premium voices and content creation. It supports multiple languages and dialects, streaming speech synthesis, free-style instruction following, and fine-grained tags to control emotion, tone, persona, speaking rate, and volume. For voice cloning, it is more robust to noise and reverberation in reference audio, improving audio quality, clarity, and expressiveness. The Plus version emphasizes audio quality and resolution for professional scenarios such as brand voices, audio content creation, film and television dubbing, and character performance.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen-audio-3.1-tts-flash
Qwen-Audio-3.1-TTS-Flash is a high-performance text-to-speech model designed for real-time interactions. It supports multiple languages and dialects, streaming speech synthesis, free-style instruction following, and fine-grained tag control for emotion, tone, persona, speaking rate, and volume. In voice cloning scenarios, it is more robust to noise and reverberation in reference audio, with improved audio quality, clarity, and expressiveness. The Flash version is optimized for low-latency applications such as voice assistants, real-time conversations, and customer service.
qwen-image-2.1-pro
Qwen-Image-2.1-Pro is an image model in the Qwen family that balances generation quality, inference efficiency, and cost. Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model, with just 7B parameters in its visual generation component and native support for generating and editing transparent images. This update introduces four key improvements: Compact and efficient: A lightweight architecture and inference optimizations balance image quality and computational cost. Native transparency, unified creation and editing: Generate regular or transparent images from prompts, edit transparent layers, and extract subjects from photos. Versatile editing: Use up to 10 reference images, specify local edits, preserve people and products, and handle a wide range of tasks. Realistic textures and refined aesthetics: Improved typography, portrait lighting, and fine details make generated images more visually compelling.
qwen-image-2.1-turbo
Qwen-Image-2.1-Turbo is an open-source image generation model in the Qwen-Image series with excellent overall performance and cost-effectiveness. Qwen-Image-2.1 offers four key strengths: Compact Model, Strong Performance, Exceptional Value. Native Transparency, Unified Generation and Editing. Versatile Editing: Multiple References, Local Control, Fidelity, and Comprehensive Capabilities. Realistic Textures, Elegant Typography, and Beautiful Portraits.
decision-model-preview
A structured decision model designed for high-frequency business decisions. It performs classification, yes-or-no decisions, and scoring in parallel based on text or business context, returning probability distributions and confidence scores for scenarios such as ticket routing, content moderation, agent routing, and result validation.
qwen3.8-omni-flash-realtime
Supports real-time audio and video interaction with text and audio output, adding multichannel audio, video aggregation, and remote MCP tools. Supports WebSocket, WebRTC, and AOQ access.
qwen-audio-3.1-asr-flash-message
Qwen-Audio-3.1-ASR-Flash-Message is a generative speech recognition model optimized for voice messages, voice input, and enterprise scenarios. It supports Chinese, English, and other languages, with optimizations for Indonesian, Japanese, Thai, Filipino, Vietnamese, Korean, and Malay. The model supports incremental streaming transcription, P0/P1 hotwords, text normalization, recognition informed by preceding transcripts and business context, and optional native text polishing. It improves recognition stability with multiple speakers, background conversations, far-field audio, and environmental noise. It supports automatic language detection or specified target languages, and can transcribe Chinese dialects as spoken or convert them to standard Mandarin text.
qwen-audio-3.1-asr-flash-streaming
Qwen-Audio-3.1-ASR-Flash-Streaming is an end-to-end real-time speech recognition model designed for scenarios such as real-time conferencing, live subtitles, and intelligent interaction. It features low latency and high accuracy in streaming speech transcription. The model supports multilingual and multi-regional Chinese dialect recognition, seamless switching between Chinese and English, hot word and context enhancement, punctuation prediction, and text normalization. It also exhibits strong noise robustness and can adapt to complex acoustic environments. Furthermore, this version supports controllable ASR/AST output for multiple dialects.
qwen-audio-3.1-asr-flash
Qwen-Audio-3.1-ASR-Flash is a large-scale model supporting efficient short speech recognition, designed for high-quality, multilingual, and cultural content transcription scenarios. The model supports multilingual and multi-regional Chinese dialect recognition, and is optimized for the rhythm, cadence, and classical Chinese expression of ancient Chinese poetry. It features context enhancement, punctuation prediction, and text normalization capabilities. This version also supports controllable ASR/AST output for multiple dialects. It includes native transcription and polishing capabilities, as well as industrial-grade multi-person dialogue role-based transcription capabilities.
qwen-audio-3.1-asr-flash-filetrans
Qwen-Audio-3.1-ASR-Flash-Filetrans is an end-to-end offline speech recognition model designed for scenarios such as meeting transcription, content production, and call analysis. It supports multilingual and multi-regional Chinese dialect recognition. The model features high-precision transcription, hot word and context enhancement, speaker separation, punctuation prediction, and text normalization capabilities. It maintains stable performance even in complex noisy environments, making it the preferred choice for long audio transcription. This version also supports controllable ASR/AST output for multiple dialects.
glm-5.3-prime
GLM-5.3-Prime is a high-speed version of GLM-5.3. It supports a 1M-token context window, up to 128K output tokens, and three thinking-depth levels. It provides coding, agent, and cybersecurity capabilities, with a 50% improvement over its predecessor on an internal coding benchmark, and is suited to end-to-end project delivery and long-running complex tasks.
qwen-audio-3.1-realtime-plus
Supports real-time duplex speech conversations using the same integration protocol as 3.0 Plus. Retains existing voices and adds eight system voices, with a 262,144-token context window, function calling, web search, and voice cloning.
happyoyster-1.0-adventure
HappyOyster Adventure is a real-time interactive, explorable open-world model. It generates world scenes from multimodal (text/image) input, enabling immersive, boundless exploration through real-time interaction, movement, and camera control.
happyoyster-1.0-directing
HappyOyster Directing is a real-time interactive, immersively performable open-world model. It generates world scenes from multimodal (text/image) input and drives real-time world performance from text at any moment.
happyoyster-1.0-acting
HappyOyster Acting is a real-time interactive role-play model. It generates characters and scenes in a specified persona style from multimodal (text/image) input, and drives real-time character performance and interaction from text at any moment.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-livetranslate-flash-realtime
The real-time version of Qwen3.8-LiveTranslate-Flash provides multilingual simultaneous translation for live audio and video. Built on Qwen-Omni with cross-language and cross-modal alignment and visual enhancement, it understands 60 languages and speaks 29 languages.
qwen3.8-max-0902
Qwen3.8-Max-0902(alias qwen3.8-max-2026-09-02)is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
qwen-mt-image-2.0
An advanced image translation engine supporting 55 source languages. It delivers pixel-perfect replication of the original layout and typography. Gain full control over professional content with custom terminology and sensitive word filtering, ensuring a flexible, accurate, and high-performance image localization service.
glm-5.3
GLM-5.3 is Z.ai's large language model, delivering enhanced coding and agentic capabilities, with a 50% improvement over its predecessor on Z.ai Code Bench and end-to-end delivery of production-grade results. It supports cybersecurity tasks such as code vulnerability discovery and delivers excellent performance on the CyberGym benchmark. With a 1M-token context window, 128K-token maximum output, and three reasoning effort levels, GLM-5.3 excels at long-horizon, complex tasks.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
qwen3.8-27b
A 27B native vision-language Dense model of the Qwen3.8 series. Compared with 3.6-27B, it focuses on improving coding and office-scenario capabilities under text and visual modalities, able to more reliably complete complex tasks end-to-end and deliver trustworthy results.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
qwen3.8-2.4t-a95b
Qwen3.8-2.4T-A95B is the open-source version of Qwen's series, released in August 2026. It adopts a sparse MoE architecture with 2.4 trillion total parameters and about 95 billion activated per step, combined with a hybrid attention mechanism, supporting a 1 million token context. Key benchmarks: GPQA Diamond 92.6, PaperBench 93.0, OSWorld 86.1, BabyVision 82.0, CodeArena global #4.
qwen-audio-3.0-realtime-plus
Qwen 3.0 Realtime Speech Model Standard Edition - next-generation duplex speech model ranked #1 globally in Artificial Analysis Speech-to-Speech benchmark. Balances model intelligence with duplex dialogue rhythm for natural interaction and enhanced response quality.
qwen-audio-3.0-realtime-flash
Qwen 3.0 Realtime Speech Dialogue Model Flash Edition - a next-generation duplex speech model ranked #1 globally in Artificial Analysis Speech-to-Speech benchmark. Combines high intelligence with optimized duplex rhythm for low-latency (parallel inference, full-streaming optimization) 'fast and smart' dialogue experience focusing on response speed.
wan3.0-video
Wan3.0 is an all-in-one video generation model that uniformly supports multiple creative functions including reference, editing, replication, and driving. It supports four-modal all-reference input, up to 30 seconds of video generation, and can parse files, web pages, and complex images. With production-grade character consistency and realistic audio and visuals, it delivers an immersive audio-visual impact.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
qwen3.7-text-embedding
Qwen3-7-Text-Embedding is a multilingual text vector model based on Qwen3.7, offering significant improvements in text retrieval, clustering, and classification over text-embedding-v4. It achieves 20% better performance in MTEB multilingual, Chinese-English, and code retrieval tasks and supports customizable vector dimensions (256-2560).
qwen-audio-3.0-asr-flash-streaming
qwen-audio-3.0-asr-flash-filetrans
qwen-audio-3.0-asr-flash
Added the Qwen-Audio-3.0-ASR-Flash-Streaming (real-time), Qwen-Audio-3.0-ASR-Flash-Filetrans (non-real-time), and Qwen-Audio-3.0-ASR-Flash (non-real-time) models: Dialect support: Supports the seven major Chinese dialect groups (Mandarin, Wu, Xiang, Gan, Hakka, Min, and Yue) and more than 20 regional accents; Classical poetry optimization: Improves recognition accuracy for classical Chinese poetry, making it suitable for education, culture, and audiobook scenarios; Text optimization: Enhances punctuation prediction and text normalization, automatically converting numbers, dates, and monetary amounts to standard formats; Multilingual expansion: Supports 30 languages, including Chinese, English, Japanese, and Korean; Hotwords and context: Supports hotwords (precompiled and on-the-fly) and context input to improve recognition accuracy for domain-specific terms.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
qwen-audio-3.0-tts-plus
Qwen-Audio-3.0-TTS-Plus is a high-performance speech synthesis model, designed for high-quality speech generation scenarios. Compared with the previous version, it supports more low-resource languages and Chinese dialects, significantly improves dialect authenticity, and enhances free-style instruction following and fine-grained tag control for more accurate control over emotion, tone, character, speaking rate, volume, and synthesis style. It is also more robust under noisy and reverberant acoustic conditions, with further improvements in audio quality, clarity, resolution, and overall expressiveness. The Plus version focuses more on synthesis quality and detailed expressiveness, making it suitable for professional scenarios with higher requirements for audio quality, naturalness, and expressiveness, such as content creation, audiobooks, film and video dubbing, brand voice design, and premium speech services.
qwen-audio-3.0-tts-flash
qwen-audio-3.0-tts-flash is a high-performance speech synthesis model, optimized for real-time interactive scenarios. Compared with the previous version, it supports more low-resource languages and Chinese dialects, improves dialect authenticity, and enhances free-style instruction following and fine-grained tag control for more flexible control over emotion, tone, character, speaking rate, volume, and expressive style. It is also more robust under noisy and reverberant acoustic conditions, with improved audio quality, clarity, and overall expressiveness. The Flash version focuses on real-time synthesis, with first-packet latency controlled within 200 ms, making it suitable for voice assistants, real-time dialogue, intelligent customer service, and other low-latency interactive applications.
glm-5.2-fast-preview
GLM-5.2-Fast-Preview is the high-speed variant of Zhipu AI's GLM-5.2, with 1M context and capabilities on par with the standard version. Inference-optimized to deliver 1.5–2× the output TPS, it fits latency-sensitive use cases such as real-time chat, multi-turn agents, and streaming code generation.
wan2.7-t2v-2026-06-12
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of June 12, 2026.
wan2.7-r2v-2026-06-12
Wan2.7 reference to video, enhanced consistency & performance. Delivering superior stability for characters, props, and scenes. Supports hybrid referencing of up to 5 mixed image/video inputs and audio timbre cloning. Together with core engine upgrades, it achieves unprecedented cinematic expressive power. This version is a snapshot as of June 12, 2026.
qwen-image-2.0-pro-2026-06-22
The Qwen-Image-2.0 series full-fledged model integrates image generation and editing; it boasts more professional text rendering capabilities with 1k token command support, more delicate and realistic textures, meticulous depiction of realistic scenes, and stronger semantic adherence. The full-fledged version possesses text rendering capabilities and realistic textures in the 2.0 series.
kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
fun-asr-flash-2026-06-15
The Bailing ASR version, updated in June 2026, fully supports the seven major dialect systems of traditional Chinese (Mandarin/Wu/Xiang/Gan/Hakka/Min/Yue) and is compatible with over 20 regional Mandarin accents. It features specific optimizations for the rhythm, meter, and literary expression characteristics of classical Chinese poetry, improving the accuracy of poetry recognition and making it suitable for cultural heritage, educational explanations, and audiobooks. Improved punctuation prediction and text normalization capabilities make the output text more consistent with written expression habits; numbers, dates, and amounts are automatically converted to standard formats, enhancing readability and professionalism. The language support has been expanded to include English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, and Slovak, totaling 30 languages. It supports contextualization capabilities and can transcribe audio up to 5 minutes in length.
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
qwen3.7-max-2026-06-08
The Max model, the largest in the Qwen3.7 series, has added visual‑modal understanding compared to the May 20 snapshot, enabling it to perceive real‑world scenes and supporting multimodal interactive hybrid agent capabilities. This version is based on a snapshot taken on June 8, 2026.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
qwen3.5-livetranslate-flash-realtime
qwen3.5-livetranslate-flash-realtime-2026-05-19
The real-time version of Qwen3.5-LiveTranslate-Flash, which is a high-precision, highly responsive, and robust multilingual simultaneous audio and video interpretation model. Leveraging Qwen3.5-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3.5-LiveTranslate-Flash offers both offline and real-time audio and video translation capabilities. It can understand 60 languages and speak 29 languages.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
wan2.7-t2v-2026-04-25
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of April 25, 2026.
wan2.7-i2v-2026-04-25
Wan2.7 image to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.This version is a snapshot as of April 25, 2026.
happyhorse-1.0-video-edit
HappyHorse-1.0-V2V supports advanced video editing through natural language instructions. It allows for local or global editing of video elements using up to 5 reference images, precisely preserving original motion dynamics to achieve superior expressiveness.
happyhorse-1.0-r2v
HappyHorse-1.0-R2V supports Reference-to-Video generation, offering enhanced stability in subject and scene referencing. Capable of processing up to 9 reference images, it precisely preserves creative intent to deliver superior performance.
qwen3.5-plus-2026-04-20
The Qwen3.5 native vision-language series Plus model has seen a substantial improvement in agentic coding capabilities compared to the February 15th snapshot. Inference speed has also been significantly enhanced, while its knowledge retention, reasoning ability, and long-context processing remain at a high level, making it well-suited for complex agent-based tasks. It is ideal for applications such as coding agents, production workflows, and high-throughput scenarios. This version is based on a snapshot taken on April 20, 2026.
qwen-image-2.0-pro
qwen-image-2.0-pro-2026-04-22
The full-featured Qwen-Image-2.0 series models integrate image generation and image editing, offering enhanced text rendering with support for 1,000-token prompts, more refined realistic textures, detailed depiction of photorealistic scenes, and stronger semantic adherence. The full-featured version delivers text rendering and realistic textures in the 2.0 series.
qwen3.6-27b
The Qwen3.6 27B native vision-language dense model builds upon the 3.5-27B version, with key improvements in agentic coding capabilities and enhanced STEM reasoning and inference skills. In the vision modality, it demonstrates significant advances in spatial intelligence, object localization, and detection, while video understanding, document OCR, and visual agent capabilities continue to improve steadily.
happyhorse-1.0-t2v
HappyHorse-1.0-T2V supports text-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends text semantics to produce high-quality videos that are fluid, natural, and rich in detail.
happyhorse-1.0-i2v
HappyHorse-1.0-I2V enables image-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends both text and image semantics to produce high-quality videos that are fluid, natural, and rich in detail.
qwen3.6-flash
qwen3.6-flash-2026-04-16
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen3.6-35b-a3b
The Qwen3.6 35B-A3B native vision-language model is built on a hybrid architecture that integrates linear attention mechanisms with a sparse mixture-of-experts framework, achieving higher inference efficiency. Compared with the 3.5-35B-A3B, this model demonstrates significantly improved agentic coding capabilities, mathematical and code reasoning abilities, spatial intelligence, as well as object localization and object detection performance.
voice-enrollment
A large-model voice replication service used in conjunction with Cosyvoice-v3. Utilizing advanced large-model technology for feature extraction, it can replicate voices without a training process. Only a very short audio clip is required to quickly generate a highly similar and natural-sounding custom voice.
qwen3.6-max-preview
The Max model, the largest in the Qwen3.6 series, is now available in a preview version. At present, only its plain-text capabilities are open for experimentation. Compared with the previously released Qwen3-Max and Qwen3.6-Plus, this model features enhanced vibe coding abilities, more efficient coding agent execution, and significantly improved front-end development skills. Additionally, its long-tail knowledge retention has been further upgraded.
glm-5.1
GLM-5.1 is a model developed by Zhipu AI, specifically designed for long-horizon tasks. It has 744 billion parameters, supports an ultra-long context of 200k tokens, and can generate up to 128k tokens in a single response. GLM-5.1 excels in logical reasoning, long-text understanding, and code generation, while balancing performance with inference efficiency. It delivers outstanding results across multiple multi-task benchmarks and is well-suited for applications such as intelligent human-computer interaction, enterprise solutions, and developer assistance.
wan2.7-videoedit
Wan2.7 video edit, supports both localized and global editing with prompt. Seamlessly replace elements using image references and replicate complex dynamic processes, including motion, special effects, and camera movements.
wan2.7-t2v
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-r2v
Wan2.7 reference to video, enhanced consistency & performance. Delivering superior stability for characters, props, and scenes. Supports hybrid referencing of up to 5 mixed image/video inputs and audio timbre cloning. Together with core engine upgrades, it achieves unprecedented cinematic expressive power.
wan2.7-i2v
Wan2.7 image to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-image-pro
Wan2.7–image-pro, supports text to image, text/image to sequential images, image editing, multi-image reference generation, and interactive editing. Delivers enhanced performance in text rendering, subject consistency, and complex instruction following.
wan2.7-image
Wan2.7 – image generation and editing, supports text to image, text/image to sequential images, image editing, multi-image reference generation, and interactive editing. Delivers enhanced performance in text rendering, subject consistency, and complex instruction following.
qwen3.6-plus
qwen3.6-plus-2026-04-02
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
qwen3.5-omni-plus-realtime
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a fully evolved version of Qwen3-Omni, it supports audio input in 60+ languages, voice output in 30+ languages, and controllable voice dialogue, WebSearch and complex FunctionCall invocation, and has intelligent semantic interruption interaction capabilities. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interactive experience.
qwen3.5-omni-plus
qwen3.5-omni-plus-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a comprehensive evolution of Qwen 3-Omni, it supports over 10 hours of audio understanding and over 400 seconds of 720P (1 FPS) audio-visual understanding and dialogue. It further expands the language range, supporting audio input in 60+ languages and speech output in 30+ languages. It also possesses powerful structured audio-visual understanding capabilities and is widely used in text creation, voice assistants, multimedia analysis, and other scenarios, providing a natural and fluent multimodal understanding and interactive experience.
qwen3.5-omni-flash-realtime
qwen3.5-omni-flash-realtime-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a fully evolved version of Qwen3-Omni, it supports audio input in 60+ languages, voice output in 30+ languages, and controllable voice dialogue, WebSearch and complex FunctionCall invocation, and has intelligent semantic interruption interaction capabilities. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interactive experience.
qwen3.5-omni-flash
qwen3.5-omni-flash-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a comprehensive evolution of Qwen 3-Omni, it supports over 10 hours of audio understanding and over 400 seconds of 720P (1 FPS) audio-visual understanding and dialogue. It further expands the language range, supporting audio input in 60+ languages and speech output in 30+ languages. It also possesses powerful structured audio-visual understanding capabilities and is widely used in text creation, voice assistants, multimedia analysis, and other scenarios, providing a natural and fluent multimodal understanding and interactive experience.
qwen3.5-omni-plus-realtime-2026-03-15
Qwen 3.5-Omni is a multimodal large model from Qwen, supporting text, image, audio, and audio-visual understanding and interaction. As a fully evolved version of Qwen3-Omni, it supports audio input in 60+ languages, voice output in 30+ languages, and controllable voice dialogue, WebSearch and complex FunctionCall invocation, and has intelligent semantic interruption interaction capabilities. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interactive experience.This version is a snapshot from March 15, 2026.
deepseek-v3.2
DeepSeek-V3.2 is the official release of a model that incorporates DeepSeek Sparse Attention—a sparse attention mechanism. It's also the first model launched by DeepSeek that integrates reasoning into tool usage, supporting both reasoning-enabled and non-reasoning tool calls.
qwen-image-2.0-pro-2026-03-03
The full-featured Qwen-Image-2.0 series models integrate image generation and image editing, offering enhanced text rendering with support for 1,000-token prompts, more refined realistic textures, detailed depiction of photorealistic scenes, and stronger semantic adherence. The full-featured version delivers text rendering and realistic textures in the 2.0 series.This version is a snapshot as of March 3, 2026.
qwen-image-2.0
qwen-image-2.0-2026-03-03
The Qwen-Image-2.0 series of accelerated models integrates image generation and image editing, offering enhanced text-rendering capabilities with support for 1,000-token prompts, more realistic textures, finely detailed photorealistic scenes, and improved semantic consistency. The accelerated version effectively strikes an optimal balance between model performance and quality.
qwen3-asr-flash-2026-02-10
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in multiple languages, while ensuring precise transcription even in complex audio environments.This version is a snapshot dated February 10, 2026.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3.5-35b-a3b
The Qwen3.5 Series 35B-A3B is a native vision-language model designed with a hybrid architecture that integrates linear attention mechanisms and a sparse mixture-of-experts model, achieving higher inference efficiency. Its overall performance is comparable to that of the Qwen3.5-27B.
qwen3.5-27b
The Qwen3.5 27B native vision-language Dense model incorporates a linear attention mechanism, delivering fast response times while balancing inference speed and performance. Its overall capabilities are comparable to those of the Qwen3.5-122B-A10B.
qwen3.5-122b-a10b
The Qwen3.5 122B-A10B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. In terms of overall performance, this model is second only to Qwen3.5-397B-A17B. Its text capabilities significantly outperform those of Qwen3-235B-2507, and its visual capabilities surpass those of Qwen3-VL-235B.
qwen3-coder-next
The new-generation code generation model in the Qwen3 series delivers performance close to that of Qwen3-Coder-Plus while offering even better capabilities. The model has been optimized with a focus on repository-level understanding, supports multi-turn tool interactions, and enhances its compatibility with agentic coding tools.
qwen3.5-plus
qwen3.5-plus-2026-02-15
The Qwen3.5 native vision-language series Plus models are built on a hybrid architecture that integrates linear attention mechanisms with sparse mixture-of-experts models, achieving higher inference efficiency. In a variety of task evaluations, the 3.5 series consistently demonstrates performance on par with state-of-the-art leading models. Compared to the 3 series, these models show a leap forward in both pure-text and multimodal capabilities.
qwen3.5-397b-a17b
The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. It delivers state-of-the-art performance comparable to leading-edge models across a wide range of tasks, including language understanding, logical reasoning, code generation, agent-based tasks, image understanding, video understanding, and graphical user interface (GUI) interactions. With its robust code-generation and agent capabilities, the model exhibits strong generalization across diverse agent.
qwen3-asr-flash-realtime-2026-02-10
The real-time version of Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in multiple languages, while ensuring precise transcription even in complex audio environments.This version is a snapshot dated February 10, 2026.
qwen3-tts-vd-2026-01-26
Qwen3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 26, 2026.
qwen3-tts-vc-2026-01-22
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated from the qwen-voice-enrollment service, and supports speech output in 11 languages using the same voice timbre. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 22, 2026.
qwen3-tts-instruct-flash
qwen3-tts-instruct-flash-2026-01-26
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. The Instruct model processes the synthesis effect through natural language, ensuring highly appropriate emotional and expressive speech in different contexts. Currently, it supports 25 timbres for both Chinese and English Instruct adjustments.
cosyvoice-v3-plus
Cloning capability: CosyVoice-v3-plus is a large voice cloning model in the CosyVoice series from Tongyi Lab. It offers superior sound quality and cloning fidelity, ideal for professional scenarios. With just 5-20 seconds of reference audio, it can rapidly generate a highly similar and natural-sounding custom voice. Synthesis capability: CosyVoice-v3-plus is a large speech synthesis model in the CosyVoice series from Tongyi Lab. It features enhanced sound quality and expressiveness, ideal for professional scenarios. The model supports real-time, streaming text-to-speech synthesis.
cosyvoice-v3-flash
Synthesis Capabilities: CosyVoice-v3-Flash is a high-performance speech synthesis model in the CosyVoice series from Tongyi Labs, offering improved naturalness, timbre, prosody, and emotional expressiveness compared to previous versions. This model supports real-time streaming text-to-speech synthesis. Cloning Capabilities: CosyVoice-v3-Flash is also a speech cloning model in the CosyVoice series from Tongyi Labs. Compared to previous versions, it improves pronunciation accuracy and timbre similarity, and adds support for more less commonly spoken languages (German, Spanish, French, Italian, Russian, Japanese). It can quickly generate highly similar and naturally sounding custom voices from just 5-20 seconds of reference audio.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
wan2.6-r2v-flash
Wan2.6 reference to video flash, faster and more cost-effective generation. Supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances.
qwen3-rerank
A text-ranking model trained on the Qwen LLM foundation performs relevance ranking for input queries and candidate documents. It supports over 100 languages and long-text inputs, and is suitable for applications such as text retrieval and RAG. Its performance is aligned with the open-source Qwen3-Rerank series models.
qwen3-max-2026-01-23
Compared with the snapshot as of September 23, 2025, the Qwen-3 series Max model in this release achieves an effective integration of thinking and non-thinking modes, resulting in a comprehensive and substantial improvement in the model's overall performance. In thinking mode, the model simultaneously supports web search, web information extraction, and a code interpreter tool, enabling it to tackle more complex and challenging problems with greater accuracy by leveraging external tools while engaging in slow, deliberative reasoning. This version is based on a snapshot taken on January 23, 2026.
qwen3-vl-flash-2026-01-22
The Qwen3 series of small-sized visual understanding models effectively integrates thinking and non-thinking modes. Compared with the snapshot taken on October 15, 2025, the overall performance of the model has improved significantly: it delivers enhanced capabilities in general visual recognition and reasoning, and shows marked improvements in recognition accuracy across various business scenarios such as security, in-store inspections, equipment monitoring, and photo-based problem solving. This version is a snapshot as of January 22, 2026.
qwen3-tts-instruct-flash-realtime
qwen3-tts-instruct-flash-realtime-2026-01-22
qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. The Instruct model processes the synthesis effect through natural language, ensuring highly appropriate emotional and expressive speech in different contexts. Currently, it supports 25 timbres for both Chinese and English Instruct adjustments. This model is equivalent to the snapshot version released on January 22, 2026.
wan2.6-i2v-flash
Wan2.6 image to video flash, faster and more cost-effective generation. Intelligent shot scheduling enables multi‑camera storytelling, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
qwen-image-edit-max
qwen-image-edit-max-2026-01-16
The Max series Qwen's image editing models delivers more stable and versatile editing capabilities: enhanced industrial design and geometric reasoning, improved character consistency, reduced offset issues, and integrated LoRA capabilities for a wider range of image editing functions.
qwen3-tts-vd-realtime-2026-01-15
Qwen3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 15, 2026.
qwen3-tts-vc-realtime-2026-01-15
Qwen 3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated from the qwen-voice-enrollment service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from January 15, 2026.
qwen-flash-character
The Qwen Role-Playing Model Series is specifically optimized for muti-language anthropomorphic interaction scenarios. It demonstrates advanced capabilities in character consistency maintenance, context-aware dialogue progression, and empathetic engagement, enabling precise personalized character embodiment. This version significantly enhances Japanese linguistic localization (including dialects and honorifics), human-like role-playing authenticity, narrative coherence control, and scenario-based cognitive intelligence.
qwen-image-plus-2026-01-09
The Qwen series of image-generation models boasts exceptional text-rendering capabilities and excels in complex text rendering as well as a wide range of generation and editing tasks. This version, a snapshot taken on January 9, 2026, is a distilled and accelerated variant of Qwen-Image-Max, enabling faster generation of high-quality images.
2025
qwen-image-max
qwen-image-max-2025-12-30
The Max series of qwen's image generation model excels across a wide range of generation tasks. Compared with the Plus series, it significantly reduces the "AI-like" feel in generated images, enhancing their realism. It delivers more lifelike material textures for human subjects, finer and more detailed natural textures, and more visually appealing text rendering.
z-image-turbo
Z-Image-Turbo is a highly efficient image-generation model that has topped the Artificial Analysis benchmark as the world's No. 1 open-source text-to-image model. With just 6 billion parameters and an 8-step inference process, it generates photo-realistic images comparable to those produced by large-scale commercial models, while excelling in bilingual Chinese–English text rendering, complex semantic understanding, and diverse thematic generation.
qwen3-vl-plus-2025-12-19
The Qwen3 series of visual understanding models effectively integrates thinking and non-thinking modes. Compared to the snapshot released on September 23, this version delivers superior performance in reasoning and analysis tasks as well as style control, while also offering lower latency and faster response speeds. This version is based on a snapshot taken on December 19, 2025.
wan2.6-r2v
Wan2.6 reference to video, supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances. Note: When a video is used as a reference, the input video is also billed. For details, see the model billing documentation.
wan2.6-t2i
Wan2.6 text to image, Upgraded visual quality, aesthetics, and instruction-following deliver precise style control, realistic portraits, long-text understanding, and broad historical/cultural IP coverage, enabling high-quality, highly expressive visual generation.
wan2.6-image
Wan2.6 Image, An all-round image generation model that supports joint text–image reasoning, multi-image creative fusion, commercial-grade consistency, aesthetic style transfer, and precise control of framing and lighting, significantly enhancing consistency, controllability, and expressiveness in image generation.
qwen-image-edit-plus-2025-12-15
The Qianwen series of Image Editing Plus models features enhanced character consistency, industrial design capabilities, and geometric reasoning abilities compared to the snapshot as of October 30. Additionally, it integrates LoRA capabilities such as lighting effects and effectively mitigates offset issues. This version is based on a snapshot taken on December 15, 2025.
qwen3-tts-vd-realtime-2025-12-16
Qwen 3-TTS-VD model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on the voices designed by the qwen3-voice-design service, and supports speech output in 11 languages using the same voice. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust its tone according to the text, demonstrating good processing capabilities for complex text synthesis. This model is a snapshot version from December 16, 2025.
qwen-voice-design
Qwen Voice-Design model is a series of voice design models from Qwen Speech Model. It only requires a simple text description to quickly design a suitable voice. When used in conjunction with the qwen3-tts-vd-realtime model, it can design and output speech in 10 languages. Furthermore, the synthesized audio can adaptively adjust its tone based on the text and has good processing capabilities for complex text synthesis.
qwen3-omni-flash-realtime
qwen3-omni-flash-realtime-2025-12-01
The real-time version of the Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-omni-flash
qwen3-omni-flash-2025-12-01
Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-livetranslate-flash
qwen3-livetranslate-flash-2025-12-01
Qwen3-LiveTranslate-Flash is a high-precision, highly responsive, and robust multilingual real-time audio and video interpretation model. Leveraging Qwen3-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3-LiveTranslate-Flash provides both offline and real-time audio and video translation capabilities. It can understand 19 languages and speak 10 languages, and also supports 8 Chinese dialects.
wan2.6-t2v
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
wan2.6-i2v
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen3-tts-vc-realtime-2025-11-27
Qwen3-TTS-Flash model is Tongyi's real-time speech synthesis model. It can perform high-fidelity real-time speech synthesis on voices replicated by the qwen3-voice-enrollment service, and supports speech output in 11 languages with the same voice timbre. This model has been trained on massive amounts of data, and the synthesized audio can adaptively adjust the tone according to the text, and it also has good processing capabilities for complex text synthesis.This model is provided as a snapshot version.
qwen3-tts-flash-realtime
qwen3-tts-flash-realtime-2025-11-27
The Qwen3-TTS-Flash-Realtime model is Tongyi's real-time speech synthesis foundation model, featuring 17 expressive voices while delivering low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on extensive datasets, the system autonomously adjusts vocal tones based on text semantics and demonstrates robust capabilities for complex content synthesis.
qwen3-tts-flash
qwen3-tts-flash-2025-11-27
The Qwen3-TTS-Flash is Tongyi's offline text-to-speech foundation model, featuring 17 expressive voices while enabling low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on massive datasets, the system automatically adjusts vocal tones based on text semantics and demonstrates robust capabilities for synthesizing complex content.
qwen-voice-enrollment
The Qwen Voice-Enrollment model is a series of voice replication models from the qwen speech model. It can quickly replicate highly similar voices using audio of only 5 seconds or more. When used in conjunction with the qwen3-tts-vc-realtime model, it can replicate a person's voice with high fidelity and output speech in 10 languages. Furthermore, the synthesized audio can adaptively adjust its tone according to the text and has good processing capabilities for complex text synthesis.
qwen-vl-ocr-2025-11-20
This model is a snapshot version from November 20, 2025, and is based on the Qwen-VL3 architecture with a comprehensive upgrade. It features significant improvements in document parsing and text localization capabilities, as well as substantial reductions in end-to-end latency and illusions.
fun-asr-2025-11-07
The Fun ASR version, updated in April 2026, fully supports the seven major dialect systems of traditional Chinese (Mandarin/Wu/Xiang/Gan/Hakka/Min/Yue) and is compatible with over 20 regional Mandarin accents. It features specific optimizations for the rhythm, meter, and literary expression characteristics of classical Chinese poetry, improving the accuracy of poetry recognition and making it suitable for cultural heritage, educational explanations, and audiobooks. Improved punctuation prediction and text normalization capabilities make the output text more consistent with written expression habits; numbers, dates, and amounts are automatically converted to standard formats, enhancing readability and professionalism. The language support has been expanded to include English, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Filipino, Hindi, Arabic, French, German, Spanish, Portuguese, Russian, Italian, Dutch, Swedish, Danish, Finnish, Norwegian, Greek, Polish, Czech, Hungarian, Romanian, Bulgarian, Croatian, and Slovak, totaling 30 languages. This is a snapshot released on November 7, 2025.
qwen-vl-ocr
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen-mt-lite
Qwen-MT-Lite is a large language model of the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 32 languages at a cost-effective price. It offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen3-asr-flash-filetrans
qwen3-asr-flash-filetrans-2025-11-17
The large file transcription version of Qwen3-ASR-Flash. Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in multiple languages, ensuring precise transcription even in complex audio environments.
fun-asr-realtime-2025-11-07
This is the real-time version of Tongyi Lab's next-generation end-to-end speech recognition model, based on leading proprietary speech technology, and boasts exceptional contextual awareness and high-precision speech transcription capabilities. Based on an end-to-end architecture, Fun-ASR integrates innovative RAG technology, supporting multi-dimensional features such as large-scale hotword customization, automatic filtering of sensitive and modal particles, ITN normalization, and punctuation prediction, significantly improving overall recognition accuracy and contextual relevance. Furthermore, Fun-ASR supports flexible switching between Chinese and English, covers multiple regional dialects, and boasts enhanced noise robustness, adapting to diverse and complex environments.This is a snapshot released on November 7, 2025.
qwen-mt-flash
Qwen-MT-Flash, a large language model from the Qwen series, has been fully upgraded with the Qwen 3 architecture for significantly enhanced performance and translation quality. It provides rapid, cost-effective translation across 92 languages, while supporting advanced features such as terminology intervention, format preservation, and domain-specific adaptation. It is the ideal choice for applications requiring a powerful balance of speed, quality, and cost.
wan2.2-animate-move
wan2.2-animate-move is a character animation generation model. Users simply upload a character photo and a reference performance video, and the model transfers the expressions and actions from the video onto the character in the image, producing a high-fidelity animated video.
wan2.2-animate-mix
wan2.2-animate-mix is a character replacement model product. By uploading a character photo and a performance video, users can accurately replace the character in the original video with the character from the photo, while completely preserving environmental details such as the scene, lighting, and color tone of the original video.
tongyi-embedding-vision-plus
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
tongyi-embedding-vision-flash
Embedding-Vision is a vision-centric multimodal embedding model powered by an LLM, featuring outstanding domain-specific performance and high cost-effectiveness in various domains (e.g., e-commerce, photo galleries, security, autonomous driving). With support for text, image, and video, it is applicable to downstream retrieval tasks, including text-to-image, image-to-image, text-to-video and video-to-video.
qwen-image-edit-plus
qwen-image-edit-plus-2025-10-30
The qwen series of image editing Plus models further optimizes inference performance and system stability based on the initial Edit model, significantly reducing the response time for image generation and editing. It also supports returning multiple images in a single request, greatly enhancing user experience.
qwen3-asr-flash-realtime
qwen3-asr-flash-realtime-2025-10-27
The real-time version of Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves highly accurate speech recognition, automatically determining the language and accurately identifying speech in 11 languages, while ensuring precise transcription even in complex audio environments.
qwen3-vl-32b-thinking
The largest dense model in the Qwen3-VL series, its reasoning version boasts multimodal reasoning capabilities second only to Qwen3-VL-235B-Thinking. It excels in STEM and math problem-solving, general image and video understanding, and achieves state-of-the-art performance in multimodal agent capabilities, making it ideal for complex multimodal reasoning tasks.
qwen3-vl-32b-instruct
The largest dense model in the Qwen3-VL series, in its non-inference version, delivers overall performance second only to Qwen3-VL-235B-Instruct. It excels in document recognition and comprehension, demonstrates strong spatial awareness and object identification capabilities, and achieves state-of-the-art performance in 2D visual detection and spatial reasoning. It is well-suited for complex perception tasks across a wide range of general-purpose scenarios.
qwen3-vl-flash
qwen3-vl-flash-2025-10-15
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-30b-a3b-thinking
The Thinking version of the second-largest MoE model in the Qwen3-VL series features fast response speeds and enhanced multimodal understanding and reasoning capabilities, visual agents, and support for extremely long contexts such as lengthy videos and documents. It also boasts comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities, making it well-suited for complex real-world tasks.
qwen3-vl-30b-a3b-instruct
The Instruct version of the second-largest MoE model in the Qwen3-VL series offers rapid response speeds and supports extremely long contexts like lengthy videos and documents. It includes comprehensively upgraded image/video understanding, spatial awareness, and object recognition capabilities, as well as 2D/3D visual localization, enabling it to handle intricate real-world challenges.
qwen3-vl-8b-thinking
The Thinking version of the 8B Dense model in the Qwen3-VL series consumes less GPU memory and is capable of performing multimodal understanding and reasoning. It supports extremely long contexts such as lengthy videos and documents, 2D/3D visual localization, and features comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities.
qwen3-vl-8b-instruct
The Instruct version of the 8B Dense model in the Qwen3-VL series requires less GPU memory and provides comprehensively upgraded image/video understanding, support for extremely long contexts like lengthy videos and documents, spatial awareness, and object recognition capabilities, making it suitable for tackling complex real-world tasks.
fun-asr-mtl
fun-asr-mtl-2025-08-25
Fun's multilingual speech recognition model supports over 31 languages and allows for free language switching, making it the top choice for users expanding overseas, especially to Southeast Asia. Fun-asr is an upgraded version of this model; switching to Fun-asr is recommended.
wan2.5-t2i-preview
The upgraded Wan2.5 Preview text to image model, newly upgraded model architecture significantly enhances visual aesthetics, design sensibility, and realistic texture. It excels in precise instruction adherence, generates text proficiently in English, Chinese, and less common languages, and supports the generation of complex structured long texts, charts, and architectural diagrams.
wan2.5-i2i-preview
The upgraded Wan2.5 Preview image edit model, newly upgraded model architecture supports rich image editing capabilities via instruction control, with enhanced instruction adherence. It also enables multi-image reference generation with high consistency and demonstrates excellent text generation performance.
qwen3-max
qwen3-max-2025-09-23
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-livetranslate-flash-realtime
The real-time version of Qwen3-LiveTranslate-Flash, which is a high-precision, highly responsive, and robust multilingual simultaneous audio and video interpretation model. Leveraging Qwen3-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3-LiveTranslate-Flash offers both offline and real-time audio and video translation capabilities. It can understand 19 languages and speak 10 languages, including 8 Chinese dialects.
fun-asr
fun-asr-2025-08-25
This is a new-generation large speech recognition model that focuses on Mandarin Chinese, English, and Japanese. It supports a wide range of regional dialects, offers enhanced noise robustness, and adapts to diverse and complex environments.This is a snapshot released on August 25, 2025.
wan2.5-t2v-preview
The upgraded Wan2.5 Preview text to video model, newly upgraded model architecture supports synchronized audio generation with visuals, enables 10-second long video generation, and offers enhanced instruction adherence, improved motion capabilities, and superior image quality.
wan2.5-i2v-preview
The upgraded Wan2.5 Preview image to video model, newly upgraded model architecture supports synchronized audio generation with visuals, enables 10-second long video generation, and offers enhanced instruction adherence, improved motion capabilities, and superior image quality.
qwen3-vl-plus
qwen3-vl-plus-2025-09-23
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen3-vl-235b-a22b-thinking
Qwen3 series VL models feature significantly enhanced multimodal reasoning capabilities, with a particular focus on optimizing the model for STEM and mathematical reasoning. Visual perception and recognition abilities have been comprehensively improved, and OCR capabilities have undergone a major upgrade.
qwen3-vl-235b-a22b-instruct
The Qwen3 series VL models has been comprehensively upgraded in areas such as visual coding and spatial perception. Its visual perception and recognition capabilities have significantly improved, supporting the understanding of ultra-long videos, and its OCR functionality has undergone a major enhancement.
qwen3-livetranslate-flash-realtime-2025-09-22
The real-time version of Qwen3-LiveTranslate-Flash, which is a high-precision, highly responsive, and robust multilingual simultaneous audio and video interpretation model. Leveraging Qwen3-Omni's powerful infrastructure, massive multimodal data, cross-language and cross-modal alignment, and visual enhancement technologies, Qwen3-LiveTranslate-Flash offers both offline and real-time audio and video translation capabilities. It can understand 19 languages and speak 10 languages, including 8 Chinese dialects.This version is a snapshot version from September 22, 2025.
qwen3-coder-plus-2025-09-23
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability. This is a snapshot from 23 September, 2025.Compared to the previous version (snapshot from July 22), it demonstrates improved robustness in downstream task performance and tool invocation, along with enhanced code security.
qwen-image-plus
The first image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
fun-asr-realtime
This is the real-time version of Tongyi Lab's next-generation end-to-end speech recognition model, based on leading proprietary speech technology, and boasts exceptional contextual awareness and high-precision speech transcription capabilities. Based on an end-to-end architecture, Fun-ASR integrates innovative RAG technology, supporting multi-dimensional features such as large-scale hotword customization, automatic filtering of sensitive and modal particles, ITN normalization, and punctuation prediction, significantly improving overall recognition accuracy and contextual relevance. Furthermore, Fun-ASR supports flexible switching between Chinese and English, covers multiple regional dialects, and boasts enhanced noise robustness, adapting to diverse and complex environments.
qwen3-omni-30b-a3b-captioner
Qwen3-Omni-30b-a3b-Captioner is a powerful fine-grained audio analysis model designed to generate accurate and comprehensive content descriptions in complex and changing audio scenarios. It can automatically parse and describe various audio content, from complex speech and ambient sounds to music and film and television sound effects, and can maintain stable and reliable output even in multi-source and mixed environments.
qwen3-omni-flash-realtime-2025-09-15
The real-time version of the Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.
qwen3-omni-flash-2025-09-15
Qwen3-Omni-Flash multimodal large-scale model, based on the Thinker–Talker Mixed Expert (MoE) architecture, supports efficient understanding and speech generation of text, images, audio, and video. It can interact with text in 119 languages and speech in 20 languages, generating human-like speech for precise cross-lingual communication. The model boasts powerful command-following and system prompt customization capabilities, flexibly adapting to conversational styles and character settings. It is widely used in scenarios such as text creation, voice assistants, and multimedia analysis, providing a natural and smooth multimodal interaction experience.This version is a snapshot version from September 15, 2025.
qwen3-tts-flash-realtime-2025-09-18
The Qwen3-TTS-Flash-Realtime-2025-09-18 model is Tongyi's real-time speech synthesis foundation model, featuring 17 expressive voices while delivering low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on extensive datasets, the system autonomously adjusts vocal tones based on text semantics and demonstrates robust capabilities for complex content synthesis.This model is provided as a snapshot version.
qwen3-tts-flash-2025-09-18
The Qwen 3-TTS-Flash-2025-09-18 is Tongyi's offline text-to-speech foundation model, featuring 17 expressive voices while enabling low-latency, high-stability audio synthesis. It supports multilingual and dialect outputs with consistent voice characteristics across languages. Trained on massive datasets, the system automatically adjusts vocal tones based on text semantics and demonstrates robust capabilities for synthesizing complex content. This model is provided as a snapshot version.
qwen3-asr-flash
qwen3-asr-flash-2025-09-08
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in 11 languages, ensuring precise transcription even in complex audio environments.
qwen-vl-plus
Qwen-VL-Plus is the enhanced version of the large visual language model. It significantly improves detail recognition and text recognition capabilities, supporting images with resolutions exceeding one million pixels and any aspect ratio specifications. The model delivers exceptional performance across a wide range of visual tasks.
qwen-vl-max
Qwen-VL-Max is a large-scale visual language model of the Qwen series. Compared to the Plus version, it further enhances visual reasoning capabilities and instruction-following abilities, offering higher levels of visual perception and cognition. It delivers optimal performance on more complex tasks.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
wan2.2-kf2v-flash
The all-new Wan2.2 First and Last Frame to video model is here. We've optimized motion stability and success rates, enhanced prompt adherence, and enabled seamless transitions between two images.
qwen3-next-80b-a3b-thinking
A new generation of Qwen3-based open-source thinking mode models. This version offers improved instruction following and streamlined summary responses over the previous iteration (Qwen3-235B-A22B-Thinking-2507).
qwen3-next-80b-a3b-instruct
A new generation of open-source, non-thinking mode model powered by Qwen3. This version demonstrates superior Chinese text understanding, augmented logical reasoning, and enhanced capabilities in text generation tasks over the previous iteration (Qwen3-235B-A22B-Instruct-2507).
qwen-plus-2025-09-11
As of the September 11, 2025 snapshot, this release features enhanced instruction following and streamlined summary responses in thinking mode. Non-thinking mode provides superior Chinese text understanding and augmented logical reasoning capabilities. The model supports a 1M context length with tiered pricing.
qwen3-max-preview
A preview version of the Max model in the Qwen 3 series, achieving an effective integration of thinking and non-thinking modes. In thinking mode, there is a significant enhancement in capabilities such as intelligent agent programming, common-sense reasoning, and reasoning across mathematics, science, and general domains.
text-embedding-v4
The General Text Vector V4 version is a multi-language text vector model developed by the Tongyi Lab based on Qwen3. Compared to the V3 version, it significantly improves performance in text retrieval, clustering, and classification tasks. It achieves a 15% to 40% improvement in evaluation tasks such as MTEB multilingual, Chinese-English, and code retrieval. Additionally, it supports user-defined vector dimensions ranging from 64 to 2048.
qwen-image-edit
The first Qwen image editing model extends Qwen-Image's text rendering to editing tasks. It offers precise bilingual (Chinese/English) text editing, dual visual and semantic editing, and strong cross-benchmark performance.
wan2.2-i2v-flash
The upgraded Wan 2.2 image to video Flash model, delivers faster speed with optimized stability, more powerful prompt following, improved consistency for text, portraits, and products, and precise shot control.
qwen-omni-turbo-realtime
qwen-omni-turbo-realtime-latest
qwen-omni-turbo-realtime-2025-05-08
The real-time version of Qwen's new large multimodal understanding and generation model, suitable for real-time audio interaction scenarios. It supports the understanding of audio accompanied by text, images, and video mixed inputs, and can simultaneously generate speech and text in stream, providing four natural tones.
qwen-image
The first image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
qwen-flash
qwen-flash-2025-07-28
The Qwen3 Flash model offers a powerful fusion of thinking and non-thinking modes with dynamic in-conversation switching, excelling in complex reasoning while showing significant gains in instruction following and text comprehension. It supports a 1M context length and is billed on a tiered model corresponding to context usage.
qwen3-coder-30b-a3b-instruct
Qwen3-based code generation model that inherits the coding agent ability of Qwen3-Coder-480B-A35B-Instruct; code capability reaches SOTA at the same scale.
qwen-plus-latest
The Qwen series model optimized for balanced performance, offering inference efficiency between Qwen-Max and Qwen-Turbo, is designed to handle moderately complex tasks effectively. This dynamically updated version implements changes without prior notice.
qwen3-30b-a3b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-30B-A3B) excels in complex thinking tasks, including logic, math, science, code, and other challenging scenarios; instruction following, text understanding, and multilingual translation capabilities significantly improved.
qwen3-coder-flash
qwen3-coder-flash-2025-07-28
Based on Qwen3, this code generation model inherits the coding agent capabilities of Qwen3-Coder-Plus and supports multi-turn tool interaction. It features focused optimizations on repository-level understanding and enhanced tool-calling stability.
qwen3-30b-a3b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-30B-A3B) shows major improvements in Chinese, English, and overall multilingual general capabilities. Optimized for subjective open-ended tasks, delivering responses significantly more aligned with user preferences and more helpful.
qwen-plus-2025-07-28
Qwen3 series Plus model, integrates thinking and non-thinking modes and can switch modes during dialogue. Compared to the prior version, adds dedicated enhancements for Chinese & English capabilities and tool calling. This is a snapshot from 28 July, 2025; first to support 1 M context length, and uses tiered pricing.
wan2.2-t2v-plus
The upgraded Wan 2.2 Plus text to video model, delivers higher quality results with stable sweeping complex movements, cinematic vision control, more powerful prompt following, and realistic world recreation.
wan2.2-t2i-plus
The upgraded Wan 2.2 Plus text to image model, delivers richer image detail with enhanced creativity, stability, and realism. It also features stronger prompt following and native support for multiple styles. Up to 2 million pixel generation and prompt enhancement are supported as well.
wan2.2-t2i-flash
The upgraded Wan 2.2 Flash text to image model, delivers faster speed with enhanced creativity, stability, and realism. It also features stronger prompt following and native support for multiple styles. Up to 2 million pixel generation and prompt enhancement are supported as well.
wan2.2-i2v-plus
The upgraded Wan 2.2 Plus image to video model, delivers higher quality results with optimized stability, more powerful prompt following, improved consistency for text, portraits, and products, and precise shot control.
qwen3-235b-a22b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-235B-A22B) shows major improvements in logical ability, general capabilities, knowledge enhancement, and creativity, suitable for high-difficulty, strong-thinking scenarios.
qwen-mt-turbo
Qwen-MT-Turbo is a large language model within the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 92 languages at a cost-effective price point. It also offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen-mt-plus
Qwen-MT-Plus, the translation model from our Qwen series, is now fully upgraded with the Qwen3 architecture. It supports 92 languages and delivers exceptionally accurate and natural-sounding translations. Its advanced capabilities in contextual understanding, terminology control, and format preservation make it a superior choice over traditional models, especially for specialized domains.
qwen3-coder-plus
qwen3-coder-plus-2025-07-22
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability.
qwen3-coder-480b-a35b-instruct
Qwen3-based code generation model with strong coding agent power; code capability reaches open-source SOTA.
qwen3-235b-a22b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-235B-A22B) shows slight improvements in subjective creativity and model safety.
qwen-plus-2025-07-14
The Plus model of the Qwen3 Series, achieving effective integration of thinking mode and non-thinking mode, allows switching modes during conversations. This is a snapshot from July 14, 2025. Compared to the previous version, there has been a significant improvement in both Chinese and English capabilities under non-thinking mode, with enhanced tool-calling abilities.
qwen-plus-character-ja
The Qwen Role-Playing Model Series is specifically optimized for Japanese anthropomorphic interaction scenarios. It demonstrates advanced capabilities in character consistency maintenance, context-aware dialogue progression, and empathetic engagement, enabling precise personalized character embodiment. This version significantly enhances Japanese linguistic localization (including dialects and honorifics), human-like role-playing authenticity, narrative coherence control, and scenario-based cognitive intelligence.
qwen-omni-turbo
qwen-omni-turbo-latest
qwen-omni-turbo-2025-03-26
The new multimodal understanding and generation model, supporting text, image, speech, and video input as well as mixed input understanding. It can simultaneously stream text and speech, with significantly enhanced speed for multimodal content understanding. Additionally, it offers four natural dialogue tones.
wan2.1-t2i-turbo
Wan2.1 Text-to-Image Turbo version, faster generation speed. Upgraded in image beauty, realism, and artistry. Stronger semantic understanding ability, rich style generalization ability, supports up to 2 million pixel generation, supports smart prompt rewriting.
wan2.1-t2i-plus
Wan2.1 Text-to-Image Plus version, Generate more image details. Upgraded in image beauty, realism, and artistry. Stronger semantic understanding ability, rich style generalization ability, supports up to 2 million pixel generation, supports smart prompt rewriting.
wan2.1-vace-plus
Wan2.1-VACE-Plus All-in-One Video Creation and Editing model.It supports local editing, video repainting, background outpainting, duration extension, image reference, and other video editing and generation tasks, and supports multimodal conditional control through text, images, and videos.
qwen3-8b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-7B.
qwen3-32b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-32B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-30b-a3b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability rivals QwQ-32B with a smaller parameter size, and its general capability significantly surpasses Qwen2.5-14B, reaching the SOTA level in the same scale industry.
qwen3-235b-a22b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-72B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-14b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-14B.
qwen-plus-2025-04-28
The Plus model of the Qwen3 series, effectively integrates thinking mode and non-thinking mode, allowing for mode switching during conversations. Its reasoning capabilities significantly surpass those of QwQ, and its general capabilities notably exceed those of Qwen2.5-Plus, reaching the SOTA level in the same scale within the industry. This model is the snapshot from April 28, 2025.
wan2.1-kf2v-plus
Wan2.1 start and end frames to video Plus version, generate a smooth transition video for two images. Support for large and complex movements, adherence to physical laws, rich artistic styles, and visual quality at the film and television level. The ability to follow instructions is further enhanced, resulting in richer details in the generated video.
qwen2.5-omni-7b
The new multi-modal understanding and generation model trained based on Qwen2.5. It supports text, image, speech, video, and mixed input understanding and can simultaneously generate streams of text and speech, significantly improves the speed of multi-modal content understanding. It provides four natural tones.
qvq-max
The Tongyi Qianwen QVQ visual reasoning model supports visual input and chain-of-thought output, demonstrating stronger capabilities in mathematics, programming, visual analysis, creation, and general tasks.
qwq-plus
The enhanced version of the Qwen QwQ reasoning model, trained on the Qwen2.5 model, has significantly improved its reasoning capabilities through reinforcement learning. The model's core metrics in mathematics and coding (e.g., AIME 24/25, LiveCodeBench) as well as some general metrics (e.g., IFEval, LiveBench) have reached the level of the full version of DeepSeek-R1.
wan2.1-i2v-turbo
Wan2.1 image to video Turbo version, make the static image generated video. Support for large and complex movements, adherence to physical laws, artistic styles, and visual quality of movies. The ability to follow instructions is further improved, and generation more fast.
qwen-turbo
The Turbo model of the Qwen3 series. It effectively integrates thinking mode and non-thinking mode, allowing for mode switching during conversations. Its reasoning capabilities rival those of QwQ-32B with a smaller parameter size, while its general capabilities significantly surpass those of Qwen2.5-Turbo, achieving the SOTA level in the same scale within the industry.
qwen-plus-2025-01-25
The Qwen series of models, which are well-balanced in capabilities, offer reasoning performance and speed that fall between Qwen-Max and Qwen-Turbo, making them suitable for moderately complex tasks. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen-max
Qwen-Max supports a parameter scale of hundreds of billions and multiple input languages such as Chinese and English. Qwen-Max is updated in a rolling manner. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
wan2.1-i2v-plus
Wan2.1 image to video Plus version, make the static image generated video. Support for large and complex movements, adherence to physical laws, artistic styles, and visual quality of movies. The ability to follow instructions is further improved, and better video quality.
wan2.1-t2v-turbo
Wan2.1 text to video Turbo version, one sentence generated video. Support for large and complex movements, adherence to physical laws, artistic styles, and visual quality of movies. The ability to follow instructions is further improved, and the generation speed is faster.
wan2.1-t2v-plus
Wan2.1 text to video Plus version, one sentence generated video. Support for large and complex movements, adherence to physical laws, artistic styles, and visual quality of movies. The ability to follow instructions is further improved, and better video quality.
2024
text-embedding-v3
The general text vectorization model is a multilingual unified text vectorization model developed by Tongyi Lab based on the large language model (LLM) foundation. This model is designed for multiple mainstream languages worldwide and provides advanced vectorization services to help developers convert text data into high-quality vector data.
US (Virginia)
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
kimi-k3.1-pro
Kimi-K3.1-Pro is a 3T-class open-weight model in Moonshot AI's K3.1 series, developed through large-scale post-training of Kimi K3. It supports a one-million-token context window, text and image inputs, text output, and native visual understanding. It is optimized for coding, 3D creation, financial research, and long-running agent tasks. It scores 65.7% on Terminal-Bench 4.0 and 64.4% on OfficeQA Pro v2, while the claim-level hallucination rate decreases from 23.4% for K3 to 8.3%.
glm-5.3
glm-5.3 is now available with the US deployment scope in the US (Virginia) region, supporting text generation, deep thinking, Function Calling, structured output, and context caching, with a 1M token context window.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen-image-2.1-pro
Qwen-Image-2.1-Pro is an image model in the Qwen family that balances generation quality, inference efficiency, and cost. Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model, with just 7B parameters in its visual generation component and native support for generating and editing transparent images. This update introduces four key improvements: Compact and efficient: A lightweight architecture and inference optimizations balance image quality and computational cost. Native transparency, unified creation and editing: Generate regular or transparent images from prompts, edit transparent layers, and extract subjects from photos. Versatile editing: Use up to 10 reference images, specify local edits, preserve people and products, and handle a wide range of tasks. Realistic textures and refined aesthetics: Improved typography, portrait lighting, and fine details make generated images more visually compelling.
qwen-image-2.1-turbo
Qwen-Image-2.1-Turbo is an open-source image generation model in the Qwen-Image series with excellent overall performance and cost-effectiveness. Qwen-Image-2.1 offers four key strengths: Compact Model, Strong Performance, Exceptional Value. Native Transparency, Unified Generation and Editing. Versatile Editing: Multiple References, Local Control, Fidelity, and Comprehensive Capabilities. Realistic Textures, Elegant Typography, and Beautiful Portraits.
qwen3.8-max
qwen3.8-max is now available with the US deployment scope in the US (Virginia) region, supporting text generation, deep thinking, visual understanding, Function Calling, structured output, and context caching, with a native 1M token context window.
qwen3.7-flash
qwen3.7-flash is now available with the US deployment scope in the US (Virginia) region, supporting text generation, deep thinking, visual understanding, Function Calling, structured output, and context caching.
qwen3.8-flash
qwen3.8-flash is now available with the US deployment scope in the US (Virginia) region, supporting text generation, deep thinking, visual understanding, Function Calling, structured output, and context caching, with a native 1M token context window and compatibility with OpenAI and Anthropic APIs.
happyoyster-1.0-adventure
HappyOyster Adventure is a real-time interactive, explorable open-world model. It generates world scenes from multimodal (text/image) input, enabling immersive, boundless exploration through real-time interaction, movement, and camera control.
happyoyster-1.0-directing
HappyOyster Directing is a real-time interactive, immersively performable open-world model. It generates world scenes from multimodal (text/image) input and drives real-time world performance from text at any moment.
happyoyster-1.0-acting
HappyOyster Acting is a real-time interactive role-play model. It generates characters and scenes in a specified persona style from multimodal (text/image) input, and drives real-time character performance and interaction from text at any moment.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-max-0902
Qwen3.8-Max-0902(alias qwen3.8-max-2026-09-02)is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
qwen-flash-character
The Qwen Role-Playing Model Series is specifically optimized for multi-language anthropomorphic interaction scenarios. It demonstrates advanced capabilities in character consistency maintenance, context-aware dialogue progression, and empathetic engagement, enabling precise personalized character embodiment.
qwen3.7-plus
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
qwen3.6-flash
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
wan2.6-i2v
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
wan2.6-t2v
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
qwen3-asr-flash
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in 11 languages, ensuring precise transcription even in complex audio environments.
glm-5.3
GLM-5.3 is Z.ai's large language model, delivering enhanced coding and agentic capabilities, with a 50% improvement over its predecessor on Z.ai Code Bench and end-to-end delivery of production-grade results. It supports cybersecurity tasks such as code vulnerability discovery and delivers excellent performance on the CyberGym benchmark. With a 1M-token context window, 128K-token maximum output, and three reasoning effort levels, GLM-5.3 excels at long-horizon, complex tasks.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
deepseek-v4-flash-0731-us
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
wan3.0-video
Wan3.0 is an all-in-one video generation model that uniformly supports multiple creative functions including reference, editing, replication, and driving. It supports four-modal all-reference input, up to 30 seconds of video generation, and can parse files, web pages, and complex images. With production-grade character consistency and realistic audio and visuals, it delivers an immersive audio-visual impact.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
glm-5.2-us
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
qwen3.6-flash-us
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
qwen3.7-plus-us
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max-us
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
qwen3.7-max-2026-06-08
The Max model, the largest in the Qwen3.7 series, has added visual‑modal understanding compared to the May 20 snapshot, enabling it to perceive real‑world scenes and supporting multimodal interactive hybrid agent capabilities. This version is based on a snapshot taken on June 8, 2026.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
deepseek-v4-pro-us
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash-us
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
kimi-k2.5
Kimi-k2.5 is a model from Moonshot AI with native multimodal architecture, supporting visual/text inputs, thinking/non-thinking modes, and both dialogue and agent tasks.
happyhorse-1.0-video-edit
HappyHorse-1.0-V2V supports advanced video editing through natural language instructions. It allows for local or global editing of video elements using up to 5 reference images, precisely preserving original motion dynamics to achieve superior expressiveness.
happyhorse-1.0-r2v
HappyHorse-1.0-R2V supports Reference-to-Video generation, offering enhanced stability in subject and scene referencing. Capable of processing up to 9 reference images, it precisely preserves creative intent to deliver superior performance.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
happyhorse-1.0-t2v
HappyHorse-1.0-T2V supports text-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends text semantics to produce high-quality videos that are fluid, natural, and rich in detail.
happyhorse-1.0-i2v
HappyHorse-1.0-I2V enables image-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends both text and image semantics to produce high-quality videos that are fluid, natural, and rich in detail.
qwen3.6-flash
qwen3.6-flash-2026-04-16
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen3.6-35b-a3b
The Qwen3.6 35B-A3B native vision-language model is built on a hybrid architecture that integrates linear attention mechanisms with a sparse mixture-of-experts framework, achieving higher inference efficiency. Compared with the 3.5-35B-A3B, this model demonstrates significantly improved agentic coding capabilities, mathematical and code reasoning abilities, spatial intelligence, as well as object localization and object detection performance.
glm-5.1
GLM-5.1 is a model developed by Zhipu AI, specifically designed for long-horizon tasks. It has 744 billion parameters, supports an ultra-long context of 200k tokens, and can generate up to 128k tokens in a single response. GLM-5.1 excels in logical reasoning, long-text understanding, and code generation, while balancing performance with inference efficiency. It delivers outstanding results across multiple multi-task benchmarks and is well-suited for applications such as intelligent human-computer interaction, enterprise solutions, and developer assistance.
qwen3.6-plus
qwen3.6-plus-2026-04-02
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
qwen-mt-lite-us
Qwen-MT-Lite is a large language model of the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 32 languages at a cost-effective price. It offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen3-vl-flash-2026-01-22-us
The Qwen3 series of small-sized visual understanding models effectively integrates thinking and non-thinking modes. Compared with the snapshot taken on October 15, 2025, the overall performance of the model has improved significantly: it delivers enhanced capabilities in general visual recognition and reasoning, and shows marked improvements in recognition accuracy across various business scenarios such as security, in-store inspections, equipment monitoring, and photo-based problem solving. This version is a snapshot as of January 22, 2026.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3.5-35b-a3b
The Qwen3.5 Series 35B-A3B is a native vision-language model designed with a hybrid architecture that integrates linear attention mechanisms and a sparse mixture-of-experts model, achieving higher inference efficiency. Its overall performance is comparable to that of the Qwen3.5-27B.
qwen3.5-27b
The Qwen3.5 27B native vision-language Dense model incorporates a linear attention mechanism, delivering fast response times while balancing inference speed and performance. Its overall capabilities are comparable to those of the Qwen3.5-122B-A10B.
qwen3.5-122b-a10b
The Qwen3.5 122B-A10B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. In terms of overall performance, this model is second only to Qwen3.5-397B-A17B. Its text capabilities significantly outperform those of Qwen3-235B-2507, and its visual capabilities surpass those of Qwen3-VL-235B.
qwen3.5-plus
qwen3.5-plus-2026-02-15
The Qwen3.5 native vision-language series Plus models are built on a hybrid architecture that integrates linear attention mechanisms with sparse mixture-of-experts models, achieving higher inference efficiency. In a variety of task evaluations, the 3.5 series consistently demonstrates performance on par with state-of-the-art leading models. Compared to the 3 series, these models show a leap forward in both pure-text and multimodal capabilities.
qwen3.5-397b-a17b
The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. It delivers state-of-the-art performance comparable to leading-edge models across a wide range of tasks, including language understanding, logical reasoning, code generation, agent-based tasks, image understanding, video understanding, and graphical user interface (GUI) interactions. With its robust code-generation and agent capabilities, the model exhibits strong generalization across diverse agent.
2025
wan2.6-r2v
Wan2.6 reference to video, supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances.
wan2.6-t2i
Wan2.6 text to image, Upgraded visual quality, aesthetics, and instruction-following deliver precise style control, realistic portraits, long-text understanding, and broad historical/cultural IP coverage, enabling high-quality, highly expressive visual generation.
wan2.6-image
Wan2.6 Image, An all-round image generation model that supports joint text–image reasoning, multi-image creative fusion, commercial-grade consistency, aesthetic style transfer, and precise control of framing and lighting, significantly enhancing consistency, controllability, and expressiveness in image generation.
wan2.6-t2v-us
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
wan2.6-i2v-us
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
wan2.6-t2v
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
wan2.6-i2v
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
qwen-plus-2025-12-01-us
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen-vl-ocr-2025-11-20
This model is a snapshot version from November 20, 2025, and is based on the Qwen-VL3 architecture with a comprehensive upgrade. It features significant improvements in document parsing and text localization capabilities, as well as substantial reductions in end-to-end latency and illusions.
qwen-mt-lite
Qwen-MT-Lite is a large language model of the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 32 languages at a cost-effective price. It offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen-mt-flash
Qwen-MT-Flash, a large language model from the Qwen series, has been fully upgraded with the Qwen 3 architecture for significantly enhanced performance and translation quality. It provides rapid, cost-effective translation across 92 languages, while supporting advanced features such as terminology intervention, format preservation, and domain-specific adaptation. It is the ideal choice for applications requiring a powerful balance of speed, quality, and cost.
qwen3-vl-32b-thinking
The largest dense model in the Qwen3-VL series, its reasoning version boasts multimodal reasoning capabilities second only to Qwen3-VL-235B-Thinking. It excels in STEM and math problem-solving, general image and video understanding, and achieves state-of-the-art performance in multimodal agent capabilities, making it ideal for complex multimodal reasoning tasks.
qwen3-vl-32b-instruct
The largest dense model in the Qwen3-VL series, in its non-inference version, delivers overall performance second only to Qwen3-VL-235B-Instruct. It excels in document recognition and comprehension, demonstrates strong spatial awareness and object identification capabilities, and achieves state-of-the-art performance in 2D visual detection and spatial reasoning. It is well-suited for complex perception tasks across a wide range of general-purpose scenarios.
qwen3-vl-flash-us
qwen3-vl-flash-2025-10-15-us
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-flash
qwen3-vl-flash-2025-10-15
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-30b-a3b-thinking
The Thinking version of the second-largest MoE model in the Qwen3-VL series features fast response speeds and enhanced multimodal understanding and reasoning capabilities, visual agents, and support for extremely long contexts such as lengthy videos and documents. It also boasts comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities, making it well-suited for complex real-world tasks.
qwen3-vl-30b-a3b-instruct
The Instruct version of the second-largest MoE model in the Qwen3-VL series offers rapid response speeds and supports extremely long contexts like lengthy videos and documents. It includes comprehensively upgraded image/video understanding, spatial awareness, and object recognition capabilities, as well as 2D/3D visual localization, enabling it to handle intricate real-world challenges.
qwen3-vl-8b-thinking
The Thinking version of the 8B Dense model in the Qwen3-VL series consumes less GPU memory and is capable of performing multimodal understanding and reasoning. It supports extremely long contexts such as lengthy videos and documents, 2D/3D visual localization, and features comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities.
qwen3-vl-8b-instruct
The Instruct version of the 8B Dense model in the Qwen3-VL series requires less GPU memory and provides comprehensively upgraded image/video understanding, support for extremely long contexts like lengthy videos and documents, spatial awareness, and object recognition capabilities, making it suitable for tackling complex real-world tasks.
qwen3-max
qwen3-max-2025-09-23
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-vl-plus
qwen3-vl-plus-2025-09-23
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen3-vl-235b-a22b-thinking
Qwen3 series VL models feature significantly enhanced multimodal reasoning capabilities, with a particular focus on optimizing the model for STEM and mathematical reasoning. Visual perception and recognition abilities have been comprehensively improved, and OCR capabilities have undergone a major upgrade.
qwen3-vl-235b-a22b-instruct
The Qwen3 series VL models has been comprehensively upgraded in areas such as visual coding and spatial perception. Its visual perception and recognition capabilities have significantly improved, supporting the understanding of ultra-long videos, and its OCR functionality has undergone a major enhancement.
qwen3-coder-plus-2025-09-23
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability. This is a snapshot from 23 September, 2025.Compared to the previous version (snapshot from July 22), it demonstrates improved robustness in downstream task performance and tool invocation, along with enhanced code security.
qwen3-asr-flash-us
qwen3-asr-flash-2025-09-08-us
Qwen3-ASR-Flash is a highly accurate, intelligent, and robust multilingual speech recognition model based on a large language model. Leveraging a powerful foundational model, massive amounts of text and multimodal data, and tens of millions of hours of audio data, Qwen3-ASR-Flash achieves high-precision speech recognition. It can automatically determine the language and accurately recognize speech in 11 languages, ensuring precise transcription even in complex audio environments.
qwen-plus-us
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen3-next-80b-a3b-thinking
A new generation of Qwen3-based open-source thinking mode models. This version offers improved instruction following and streamlined summary responses over the previous iteration (Qwen3-235B-A22B-Thinking-2507).
qwen3-next-80b-a3b-instruct
A new generation of open-source, non-thinking mode model powered by Qwen3. This version demonstrates superior Chinese text understanding, augmented logical reasoning, and enhanced capabilities in text generation tasks over the previous iteration (Qwen3-235B-A22B-Instruct-2507).
qwen-plus-2025-09-11
As of the September 11, 2025 snapshot, this release features enhanced instruction following and streamlined summary responses in thinking mode. Non-thinking mode provides superior Chinese text understanding and augmented logical reasoning capabilities. The model supports a 1M context length with tiered pricing.
qwen3-max-preview
A preview version of the Max model in the Qwen 3 series, achieving an effective integration of thinking and non-thinking modes. In thinking mode, there is a significant enhancement in capabilities such as intelligent agent programming, common-sense reasoning, and reasoning across mathematics, science, and general domains.
qwen-flash-us
qwen-flash-2025-07-28-us
The Qwen3 Flash model offers a powerful fusion of thinking and non-thinking modes with dynamic in-conversation switching, excelling in complex reasoning while showing significant gains in instruction following and text comprehension. It supports a 1M context length and is billed on a tiered model corresponding to context usage.
qwen-flash
qwen-flash-2025-07-28
The Qwen3 Flash model offers a powerful fusion of thinking and non-thinking modes with dynamic in-conversation switching, excelling in complex reasoning while showing significant gains in instruction following and text comprehension. It supports a 1M context length and is billed on a tiered model corresponding to context usage.
qwen3-coder-30b-a3b-instruct
Qwen3-based code generation model that inherits the coding agent ability of Qwen3-Coder-480B-A35B-Instruct; code capability reaches SOTA at the same scale.
qwen3-30b-a3b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-30B-A3B) excels in complex thinking tasks, including logic, math, science, code, and other challenging scenarios; instruction following, text understanding, and multilingual translation capabilities significantly improved.
qwen3-coder-flash
qwen3-coder-flash-2025-07-28
Based on Qwen3, this code generation model inherits the coding agent capabilities of Qwen3-Coder-Plus and supports multi-turn tool interaction. It features focused optimizations on repository-level understanding and enhanced tool-calling stability.
qwen3-30b-a3b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-30B-A3B) shows major improvements in Chinese, English, and overall multilingual general capabilities. Optimized for subjective open-ended tasks, delivering responses significantly more aligned with user preferences and more helpful.
qwen-plus-2025-07-28
Qwen3 series Plus model, integrates thinking and non-thinking modes and can switch modes during dialogue. Compared to the prior version, adds dedicated enhancements for Chinese & English capabilities and tool calling. This is a snapshot from 28 July, 2025; first to support 1 M context length, and uses tiered pricing.
qwen3-235b-a22b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-235B-A22B) shows major improvements in logical ability, general capabilities, knowledge enhancement, and creativity, suitable for high-difficulty, strong-thinking scenarios.
qwen-mt-plus
Qwen-MT-Plus, the translation model from our Qwen series, is now fully upgraded with the Qwen3 architecture. It supports 92 languages and delivers exceptionally accurate and natural-sounding translations. Its advanced capabilities in contextual understanding, terminology control, and format preservation make it a superior choice over traditional models, especially for specialized domains.
qwen3-coder-plus
qwen3-coder-plus-2025-07-22
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability.
qwen3-coder-480b-a35b-instruct
Qwen3-based code generation model with strong coding agent power; code capability reaches open-source SOTA.
qwen3-235b-a22b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-235B-A22B) shows slight improvements in subjective creativity and model safety.
qwen-vl-ocr
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen3-8b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-7B.
qwen3-32b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-32B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-30b-a3b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability rivals QwQ-32B with a smaller parameter size, and its general capability significantly surpasses Qwen2.5-14B, reaching the SOTA level in the same scale industry.
qwen3-235b-a22b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-72B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-14b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-14B.
Germany (Frankfurt)
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-max-0902
Qwen3.8-Max-0902(alias qwen3.8-max-2026-09-02)is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
glm-5.3
GLM-5.3 is Z.ai's large language model, delivering enhanced coding and agentic capabilities, with a 50% improvement over its predecessor on Z.ai Code Bench and end-to-end delivery of production-grade results. It supports cybersecurity tasks such as code vulnerability discovery and delivers excellent performance on the CyberGym benchmark. With a 1M-token context window, 128K-token maximum output, and three reasoning effort levels, GLM-5.3 excels at long-horizon, complex tasks.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
wan3.0-video
Wan3.0 is an all-in-one video generation model that uniformly supports multiple creative functions including reference, editing, replication, and driving. It supports four-modal all-reference input, up to 30 seconds of video generation, and can parse files, web pages, and complex images. With production-grade character consistency and realistic audio and visuals, it delivers an immersive audio-visual impact.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
qwen3.7-max-2026-06-08
The Max model, the largest in the Qwen3.7 series, has added visual‑modal understanding compared to the May 20 snapshot, enabling it to perceive real‑world scenes and supporting multimodal interactive hybrid agent capabilities. This version is based on a snapshot taken on June 8, 2026.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
kimi-k2.5
Kimi-k2.5 is a model from Moonshot AI with native multimodal architecture, supporting visual/text inputs, thinking/non-thinking modes, and both dialogue and agent tasks.
happyhorse-1.0-video-edit
HappyHorse-1.0-V2V supports advanced video editing through natural language instructions. It allows for local or global editing of video elements using up to 5 reference images, precisely preserving original motion dynamics to achieve superior expressiveness.
happyhorse-1.0-r2v
HappyHorse-1.0-R2V supports Reference-to-Video generation, offering enhanced stability in subject and scene referencing. Capable of processing up to 9 reference images, it precisely preserves creative intent to deliver superior performance.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
happyhorse-1.0-t2v
HappyHorse-1.0-T2V supports text-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends text semantics to produce high-quality videos that are fluid, natural, and rich in detail.
happyhorse-1.0-i2v
HappyHorse-1.0-I2V enables image-to-video generation, featuring highly realistic dynamic rendering. It accurately comprehends both text and image semantics to produce high-quality videos that are fluid, natural, and rich in detail.
qwen3.6-flash
qwen3.6-flash-2026-04-16
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen3.6-35b-a3b
The Qwen3.6 35B-A3B native vision-language model is built on a hybrid architecture that integrates linear attention mechanisms with a sparse mixture-of-experts framework, achieving higher inference efficiency. Compared with the 3.5-35B-A3B, this model demonstrates significantly improved agentic coding capabilities, mathematical and code reasoning abilities, spatial intelligence, as well as object localization and object detection performance.
glm-5.1
GLM-5.1 is a model developed by Zhipu AI, specifically designed for long-horizon tasks. It has 744 billion parameters, supports an ultra-long context of 200k tokens, and can generate up to 128k tokens in a single response. GLM-5.1 excels in logical reasoning, long-text understanding, and code generation, while balancing performance with inference efficiency. It delivers outstanding results across multiple multi-task benchmarks and is well-suited for applications such as intelligent human-computer interaction, enterprise solutions, and developer assistance.
qwen3.6-plus
qwen3.6-plus-2026-04-02
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3.5-35b-a3b
The Qwen3.5 Series 35B-A3B is a native vision-language model designed with a hybrid architecture that integrates linear attention mechanisms and a sparse mixture-of-experts model, achieving higher inference efficiency. Its overall performance is comparable to that of the Qwen3.5-27B.
qwen3.5-27b
The Qwen3.5 27B native vision-language Dense model incorporates a linear attention mechanism, delivering fast response times while balancing inference speed and performance. Its overall capabilities are comparable to those of the Qwen3.5-122B-A10B.
qwen3.5-122b-a10b
The Qwen3.5 122B-A10B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. In terms of overall performance, this model is second only to Qwen3.5-397B-A17B. Its text capabilities significantly outperform those of Qwen3-235B-2507, and its visual capabilities surpass those of Qwen3-VL-235B.
qwen3-coder-next
The new-generation code generation model in the Qwen3 series delivers performance close to that of Qwen3-Coder-Plus while offering even better capabilities. The model has been optimized with a focus on repository-level understanding, supports multi-turn tool interactions, and enhances its compatibility with agentic coding tools.
qwen3.5-plus
qwen3.5-plus-2026-02-15
The Qwen3.5 native vision-language series Plus models are built on a hybrid architecture that integrates linear attention mechanisms with sparse mixture-of-experts models, achieving higher inference efficiency. In a variety of task evaluations, the 3.5 series consistently demonstrates performance on par with state-of-the-art leading models. Compared to the 3 series, these models show a leap forward in both pure-text and multimodal capabilities.
qwen3.5-397b-a17b
The Qwen3.5 series 397B-A17B native vision-language model is built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. It delivers state-of-the-art performance comparable to leading-edge models across a wide range of tasks, including language understanding, logical reasoning, code generation, agent-based tasks, image understanding, video understanding, and graphical user interface (GUI) interactions. With its robust code-generation and agent capabilities, the model exhibits strong generalization across diverse agent.
qwen3-max-2026-01-23
Compared with the snapshot as of September 23, 2025, the Qwen-3 series Max model in this release achieves an effective integration of thinking and non-thinking modes, resulting in a comprehensive and substantial improvement in the model's overall performance. In thinking mode, the model simultaneously supports web search, web information extraction, and a code interpreter tool, enabling it to tackle more complex and challenging problems with greater accuracy by leveraging external tools while engaging in slow, deliberative reasoning. This version is based on a snapshot taken on January 23, 2026.
qwen3-vl-flash-2026-01-22
The Qwen3 series of small-sized visual understanding models effectively integrates thinking and non-thinking modes. Compared with the snapshot taken on October 15, 2025, the overall performance of the model has improved significantly: it delivers enhanced capabilities in general visual recognition and reasoning, and shows marked improvements in recognition accuracy across various business scenarios such as security, in-store inspections, equipment monitoring, and photo-based problem solving. This version is a snapshot as of January 22, 2026.
2025
wan2.6-r2v
Wan2.6 reference to video, supports using a specified person or any object as a reference, precisely maintaining consistency of appearance and voice, and allows multi‑character reference for joint performances.
wan2.6-t2i
Wan2.6 text to image, Upgraded visual quality, aesthetics, and instruction-following deliver precise style control, realistic portraits, long-text understanding, and broad historical/cultural IP coverage, enabling high-quality, highly expressive visual generation.
wan2.6-image
Wan2.6 Image, An all-round image generation model that supports joint text–image reasoning, multi-image creative fusion, commercial-grade consistency, aesthetic style transfer, and precise control of framing and lighting, significantly enhancing consistency, controllability, and expressiveness in image generation.
wan2.6-t2v
Wan2.6 text to video, Intelligent shot scheduling supports multi-shot storytelling, generating multi-shot narrative videos with consistent subjects, scenes, and atmosphere, with a maximum duration of 15 seconds. It delivers better instruction following, and improved visual fidelity.
wan2.6-i2v
Wan2.6 image to video, intelligent shot scheduling enables multi‑camera storytelling, delivers higher‑quality voice generation, supports stable multi‑speaker dialogue with more natural and realistic vocal timbres, and supports generation of clips up to 15 seconds in length.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen-vl-ocr-2025-11-20
This model is a snapshot version from November 20, 2025, and is based on the Qwen-VL3 architecture with a comprehensive upgrade. It features significant improvements in document parsing and text localization capabilities, as well as substantial reductions in end-to-end latency and illusions.
qwen-vl-ocr
Qwen-VL_OCR is an OCR model trained based on Qwen-VL. It aggregates various image-text recognition, parsing, and processing tasks through a unified model approach, offering powerful image-text recognition capabilities.
qwen-mt-lite
Qwen-MT-Lite is a large language model of the Qwen model series that specializes in multi-lingual translation. It provides high-quality and rapid translation services across 32 languages at a cost-effective price. It offers features such as terminology intervention, format preservation, and domain-specific translation to cater to the diverse needs of various applications, ensuring both efficiency and performance.
qwen-mt-flash
Qwen-MT-Flash, a large language model from the Qwen series, has been fully upgraded with the Qwen 3 architecture for significantly enhanced performance and translation quality. It provides rapid, cost-effective translation across 92 languages, while supporting advanced features such as terminology intervention, format preservation, and domain-specific adaptation. It is the ideal choice for applications requiring a powerful balance of speed, quality, and cost.
qwen3-vl-32b-thinking
The largest dense model in the Qwen3-VL series, its reasoning version boasts multimodal reasoning capabilities second only to Qwen3-VL-235B-Thinking. It excels in STEM and math problem-solving, general image and video understanding, and achieves state-of-the-art performance in multimodal agent capabilities, making it ideal for complex multimodal reasoning tasks.
qwen3-vl-32b-instruct
The largest dense model in the Qwen3-VL series, in its non-inference version, delivers overall performance second only to Qwen3-VL-235B-Instruct. It excels in document recognition and comprehension, demonstrates strong spatial awareness and object identification capabilities, and achieves state-of-the-art performance in 2D visual detection and spatial reasoning. It is well-suited for complex perception tasks across a wide range of general-purpose scenarios.
qwen3-vl-flash
qwen3-vl-flash-2025-10-15
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-flash
qwen3-vl-flash-2025-10-15
The Qwen3 series of small-scale visual understanding models effectively integrates thinking and non-thinking modes, delivering superior performance compared to the open-source Qwen3-VL-30B-A3B while maintaining fast response speeds. It features a comprehensive upgrade in image/video understanding, supporting ultra-long contexts such as extended videos and documents, spatial awareness, and object recognition across various domains. Equipped with 2D/3D visual localization capabilities, it is well-suited for tackling complex real-world tasks.
qwen3-vl-30b-a3b-thinking
The Thinking version of the second-largest MoE model in the Qwen3-VL series features fast response speeds and enhanced multimodal understanding and reasoning capabilities, visual agents, and support for extremely long contexts such as lengthy videos and documents. It also boasts comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities, making it well-suited for complex real-world tasks.
qwen3-vl-30b-a3b-instruct
The Instruct version of the second-largest MoE model in the Qwen3-VL series offers rapid response speeds and supports extremely long contexts like lengthy videos and documents. It includes comprehensively upgraded image/video understanding, spatial awareness, and object recognition capabilities, as well as 2D/3D visual localization, enabling it to handle intricate real-world challenges.
qwen3-vl-8b-thinking
The Thinking version of the 8B Dense model in the Qwen3-VL series consumes less GPU memory and is capable of performing multimodal understanding and reasoning. It supports extremely long contexts such as lengthy videos and documents, 2D/3D visual localization, and features comprehensively upgraded image/video understanding, spatial awareness, and object recognition abilities.
qwen3-vl-8b-instruct
The Instruct version of the 8B Dense model in the Qwen3-VL series requires less GPU memory and provides comprehensively upgraded image/video understanding, support for extremely long contexts like lengthy videos and documents, spatial awareness, and object recognition capabilities, making it suitable for tackling complex real-world tasks.
qwen3-max
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-max
qwen3-max-2025-09-23
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-vl-plus
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen3-vl-plus
qwen3-vl-plus-2025-09-23
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen3-vl-235b-a22b-thinking
Qwen3 series VL models feature significantly enhanced multimodal reasoning capabilities, with a particular focus on optimizing the model for STEM and mathematical reasoning. Visual perception and recognition abilities have been comprehensively improved, and OCR capabilities have undergone a major upgrade.
qwen3-vl-235b-a22b-instruct
The Qwen3 series VL models has been comprehensively upgraded in areas such as visual coding and spatial perception. Its visual perception and recognition capabilities have significantly improved, supporting the understanding of ultra-long videos, and its OCR functionality has undergone a major enhancement.
qwen3-coder-plus-2025-09-23
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability. This is a snapshot from 23 September, 2025.Compared to the previous version (snapshot from July 22), it demonstrates improved robustness in downstream task performance and tool invocation, along with enhanced code security.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
qwen3-next-80b-a3b-thinking
A new generation of Qwen3-based open-source thinking mode models. This version offers improved instruction following and streamlined summary responses over the previous iteration (Qwen3-235B-A22B-Thinking-2507).
qwen3-next-80b-a3b-instruct
A new generation of open-source, non-thinking mode model powered by Qwen3. This version demonstrates superior Chinese text understanding, augmented logical reasoning, and enhanced capabilities in text generation tasks over the previous iteration (Qwen3-235B-A22B-Instruct-2507).
qwen-plus-2025-09-11
As of the September 11, 2025 snapshot, this release features enhanced instruction following and streamlined summary responses in thinking mode. Non-thinking mode provides superior Chinese text understanding and augmented logical reasoning capabilities. The model supports a 1M context length with tiered pricing.
qwen3-max-preview
A preview version of the Max model in the Qwen 3 series, achieving an effective integration of thinking and non-thinking modes. In thinking mode, there is a significant enhancement in capabilities such as intelligent agent programming, common-sense reasoning, and reasoning across mathematics, science, and general domains.
qwen-flash
qwen-flash-2025-07-28
The Qwen3 Flash model offers a powerful fusion of thinking and non-thinking modes with dynamic in-conversation switching, excelling in complex reasoning while showing significant gains in instruction following and text comprehension. It supports a 1M context length and is billed on a tiered model corresponding to context usage.
qwen3-coder-30b-a3b-instruct
Qwen3-based code generation model that inherits the coding agent ability of Qwen3-Coder-480B-A35B-Instruct; code capability reaches SOTA at the same scale.
qwen3-30b-a3b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-30B-A3B) excels in complex thinking tasks, including logic, math, science, code, and other challenging scenarios; instruction following, text understanding, and multilingual translation capabilities significantly improved.
qwen3-coder-flash
qwen3-coder-flash-2025-07-28
Based on Qwen3, this code generation model inherits the coding agent capabilities of Qwen3-Coder-Plus and supports multi-turn tool interaction. It features focused optimizations on repository-level understanding and enhanced tool-calling stability.
qwen3-30b-a3b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-30B-A3B) shows major improvements in Chinese, English, and overall multilingual general capabilities. Optimized for subjective open-ended tasks, delivering responses significantly more aligned with user preferences and more helpful.
qwen-plus-2025-07-28
Qwen3 series Plus model, integrates thinking and non-thinking modes and can switch modes during dialogue. Compared to the prior version, adds dedicated enhancements for Chinese & English capabilities and tool calling. This is a snapshot from 28 July, 2025; first to support 1 M context length, and uses tiered pricing.
qwen3-235b-a22b-thinking-2507
Open-source Qwen3 thinking model; compared to the previous version (Qwen3-235B-A22B) shows major improvements in logical ability, general capabilities, knowledge enhancement, and creativity, suitable for high-difficulty, strong-thinking scenarios.
qwen-mt-plus
Qwen-MT-Plus, the translation model from our Qwen series, is now fully upgraded with the Qwen3 architecture. It supports 92 languages and delivers exceptionally accurate and natural-sounding translations. Its advanced capabilities in contextual understanding, terminology control, and format preservation make it a superior choice over traditional models, especially for specialized domains.
qwen3-coder-plus
qwen3-coder-plus-2025-07-22
Qwen3-based code generation model with strong coding agent power, excels at tool calling and environment interaction, capable of autonomous programming with outstanding code capability while maintaining general ability.
qwen3-coder-480b-a35b-instruct
Qwen3-based code generation model with strong coding agent power; code capability reaches open-source SOTA.
qwen3-235b-a22b-instruct-2507
Open-source Qwen3 non-thinking model; compared to the previous version (Qwen3-235B-A22B) shows slight improvements in subjective creativity and model safety.
qwen3-8b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability reaches the SOTA level in the same scale industry, and its general capability significantly surpasses Qwen2.5-7B.
qwen3-32b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-32B-Instruct, reaching the SOTA level in the same scale industry.
qwen3-30b-a3b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability rivals QwQ-32B with a smaller parameter size, and its general capability significantly surpasses Qwen2.5-14B, reaching the SOTA level in the same scale industry.
qwen3-235b-a22b
Achieves effective integration of thinking and non-thinking modes, allowing mode switching during conversations. Its reasoning capability significantly surpasses QwQ, and its general capability markedly exceeds Qwen2.5-72B-Instruct, reaching the SOTA level in the same scale industry.
Japan (Tokyo)
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
kimi-k3.1-pro
Kimi-K3.1-Pro is a 3T-class open-weight model in Moonshot AI's K3.1 series, developed through large-scale post-training of Kimi K3. It supports a one-million-token context window, text and image inputs, text output, and native visual understanding. It is optimized for coding, 3D creation, financial research, and long-running agent tasks. It scores 65.7% on Terminal-Bench 4.0 and 64.4% on OfficeQA Pro v2, while the claim-level hallucination rate decreases from 23.4% for K3 to 8.3%.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen-image-2.1-pro
Qwen-Image-2.1-Pro is an image model in the Qwen family that balances generation quality, inference efficiency, and cost. Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model, with just 7B parameters in its visual generation component and native support for generating and editing transparent images. This update introduces four key improvements: Compact and efficient: A lightweight architecture and inference optimizations balance image quality and computational cost. Native transparency, unified creation and editing: Generate regular or transparent images from prompts, edit transparent layers, and extract subjects from photos. Versatile editing: Use up to 10 reference images, specify local edits, preserve people and products, and handle a wide range of tasks. Realistic textures and refined aesthetics: Improved typography, portrait lighting, and fine details make generated images more visually compelling.
qwen-image-2.1-turbo
Qwen-Image-2.1-Turbo is an open-source image generation model in the Qwen-Image series with excellent overall performance and cost-effectiveness. Qwen-Image-2.1 offers four key strengths: Compact Model, Strong Performance, Exceptional Value. Native Transparency, Unified Generation and Editing. Versatile Editing: Multiple References, Local Control, Fidelity, and Comprehensive Capabilities. Realistic Textures, Elegant Typography, and Beautiful Portraits.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-max-0902
Qwen3.8-Max-0902(alias qwen3.8-max-2026-09-02)is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
glm-5.3
GLM-5.3 is Z.ai's large language model, delivering enhanced coding and agentic capabilities, with a 50% improvement over its predecessor on Z.ai Code Bench and end-to-end delivery of production-grade results. It supports cybersecurity tasks such as code vulnerability discovery and delivers excellent performance on the CyberGym benchmark. With a 1M-token context window, 128K-token maximum output, and three reasoning effort levels, GLM-5.3 excels at long-horizon, complex tasks.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
wan3.0-video
Wan3.0 is an all-in-one video generation model that uniformly supports multiple creative functions including reference, editing, replication, and driving. It supports four-modal all-reference input, up to 30 seconds of video generation, and can parse files, web pages, and complex images. With production-grade character consistency and realistic audio and visuals, it delivers an immersive audio-visual impact.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
happyhorse-1.1-t2v
HappyHorse-1.1-T2V supports Text-to-Video generation with improved semantic understanding, cinematic shot control, and dynamic motion rendering. It more accurately captures creative intent, producing high-quality videos with smoother motion, richer details, stronger visual consistency, and more natural character actions, scene atmosphere, and physical dynamics.
happyhorse-1.1-r2v
HappyHorse-1.1-R2V supports Reference-to-Video generation with significantly improved stability in subject, scene style, and visual consistency. With support for up to 9 reference images, it can more accurately understand and preserve creative intent, delivering stronger controllability and expressiveness across characters, scenes, styles, and cinematic motion.
happyhorse-1.1-i2v
HappyHorse-1.1-I2V supports Image-to-Video generation with improved visual quality, dynamic performance, and cross-clip consistency. It more accurately understands the input image and preserves creative intent, delivering significant improvements in skin texture realism, character ID consistency across clips, motion smoothness, text rendering stability, and audio-visual synchronization, producing high-quality videos with greater realism, richer details, and stronger overall consistency.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-pro
A MoE large model with 1.6 trillion parameters and 49 billion activated parameters, natively supporting context lengths of up to one million tokens. Trained on a vast corpus of high-quality data, it excels in advanced mathematical reasoning, complex logical inference, specialized coding, and deep analysis of long-form text, making it well-suited for demanding applications such as cutting-edge research, sophisticated office workflows, and advanced AI agents.
deepseek-v4-flash
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
kimi-k2.5
Kimi-k2.5 is a model from Moonshot AI with native multimodal architecture, supporting visual/text inputs, thinking/non-thinking modes, and both dialogue and agent tasks.
happyhorse-1.0-video-edit
HappyHorse-1.0-V2V supports advanced video editing through natural language instructions. It allows for local or global editing of video elements using up to 5 reference images, precisely preserving original motion dynamics to achieve superior expressiveness.
qwen3.6-flash
qwen3.6-flash-2026-04-16
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
glm-5.1
GLM-5.1 is a model developed by Zhipu AI, specifically designed for long-horizon tasks. It has 744 billion parameters, supports an ultra-long context of 200k tokens, and can generate up to 128k tokens in a single response. GLM-5.1 excels in logical reasoning, long-text understanding, and code generation, while balancing performance with inference efficiency. It delivers outstanding results across multiple multi-task benchmarks and is well-suited for applications such as intelligent human-computer interaction, enterprise solutions, and developer assistance.
wan2.7-videoedit
Wan2.7 video edit, supports both localized and global editing with prompt. Seamlessly replace elements using image references and replicate complex dynamic processes, including motion, special effects, and camera movements.
wan2.7-t2v
Wan2.7 text to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-r2v
Wan2.7 reference to video, enhanced consistency & performance. Delivering superior stability for characters, props, and scenes. Supports hybrid referencing of up to 5 mixed image/video inputs and audio timbre cloning. Together with core engine upgrades, it achieves unprecedented cinematic expressive power.
wan2.7-i2v
Wan2.7 image to video, performance fully reimagined. Delivering nuanced and organic emotional depth in narrative arcs and visceral, bone-crunching impact in action sequences. Enhanced by rhythmic cinematic cuts for unparalleled storytelling power.
wan2.7-image
Wan2.7 – image generation and editing, supports text to image, text/image to sequential images, image editing, multi-image reference generation, and interactive editing. Delivers enhanced performance in text rendering, subject consistency, and complex instruction following.
qwen3.6-plus
qwen3.6-plus-2026-04-02
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
Hong Kong (China)
2026
qwen4-max
Qwen4-Max is the next-generation flagship multimodal model in the Qwen family, combining text generation, image and video understanding, and deep reasoning. It is designed for coding, office productivity, frontend design, and complex agentic tasks. With support for web search, tool calling, and structured outputs, along with compatibility with the OpenAI and Anthropic APIs, it enables developers to build intelligent applications capable of multimodal understanding and task execution.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
deepseek-v4.1-flash-1017
DeepSeek-V4.1-Flash-1017 is an iteration of V4.1-Flash, trained with self-feedback-based reinforcement learning to deliver substantial gains over its predecessor across benchmarks such as Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench, while shortening task runtime and output length to reach a new performance–cost Pareto optimum across multiple harnesses. Its ability to build complete projects from scratch is further strengthened, enabling "wish-based" programming tasks such as game development and 3D modeling, while retaining the same 552B-parameter MoE, native multimodal visual understanding, 1M-token context window, and 384K-token maximum output.
qwen3.8-omni-flash
Qwen3.8-Omni-Flash supports text, image, audio, and video inputs with text output. It supports thinking and non-thinking modes, function calling, web search, and context caching for audio and video understanding, multimodal analysis, and agent applications.
deepseek-v4.1-flash
DeepSeek-V4.1-Flash is the lightweight model in DeepSeek's new architecture family, packing 552B total MoE parameters to deliver performance exceeding DeepSeek-V4-Pro across key benchmarks. It adopts a Causal Encoder-Decoder asymmetric design with 8B active parameters for input and 16B for output, and offers native multimodal visual understanding. KV Cache usage is cut to one-quarter of the previous generation's HBM and one-eighth of its SSD storage, dramatically lowering costs for long-context and agentic workloads. With a 1M-token context window and up to 384K-token maximum output, it balances high throughput, low latency, and exceptional cost efficiency.
qwen3.8-max-0902
Qwen3.8-Max-0902 (alias qwen3.8-max-2026-09-02) is an upgraded snapshot of qwen3.8-max. Coding capability breaks new ground, handling more complex engineering-scale projects and long-horizon autonomous development. Collaborative agent performance is significantly enhanced, with greater composure in multi-tool orchestration and end-to-end task delivery. Native vision understanding is refined across chart reasoning, document parsing, and multimodal perception — sharper and more reliable. Retains the 1M context window, thinking mode, and full tool ecosystem, evolving at a higher level of intelligence.
qwen3.8-flash
Qwen3.8-Flash is a multimodal model from the Qwen family, combining powerful reasoning and generation with remarkable speed. It natively supports a million-token context window, allowing it to process lengthy documents, entire codebases, and complex conversations in a single pass. It shines in coding assistance, agentic workflows, and visual understanding — whether it's fixing code autonomously, operating desktop applications, or analyzing charts and long videos. Fully compatible with both OpenAI and Anthropic API protocols, it integrates seamlessly with popular developer tools like Claude Code and Codex, making it easy to build high-concurrency applications and intelligent workflows. With strong performance and highly competitive inference costs, Qwen3.8-Flash is an ideal choice for developers and businesses seeking the best of both worlds in AI applications.
wan3.0-video-prime
Wan3.0-Video-Prime is the high-speed version of the Wan3.0 video generation model, with capabilities aligned with the Wan3.0-Video standard version. It supports four-modal all-reference input and can generate videos up to 30 seconds long, delivering an immersive audio-visual experience with significantly improved end-to-end speed.
kimi-k3
Kimi K3 is Kimi's model with 2.8 trillion parameters. Built on KDA hybrid linear attention (Kimi Delta Attention) and Attention Residuals technologies, it natively supports vision understanding and features a 1 million token context window. It is the world's first open-source 3-trillion-parameter model designed for long-context programming, knowledge work, and advanced reasoning.
deepseek-v4-pro-0813
A Mixture-of-Experts (MoE) large language model with 1.6 trillion total parameters and 49 billion activated parameters. It natively supports context windows of up to 1 million tokens. Trained on extensive high-quality data, the model delivers strong performance in mathematical and logical reasoning, complex reasoning, professional code generation, and in-depth long-document analysis, and is suitable for demanding scenarios such as advanced scientific research, complex enterprise workflows, and sophisticated agentic applications.
qwen-image-3.0-pro
Rich content: Supports input of up to 4.5k tokens and dense information layout with images-within-images, enabling complex layouts like newspapers, storyboards, menus, and exam papers to be generated in a single pass. Authentic detail: Supports precise rendering of text as small as 10px, and vividly reproduces fine details such as micro-expressions, pores, and individual strands of hair—approaching the quality of real photography. Deep knowledge: Supports native rendering of 12 languages and 20+ fonts, realistic simulation of mainstream interfaces such as web pages, games, and live streams, fully incorporating external knowledge. Qwen-Image-3.0-Pro isn't just pursuing "good looks"—it's pursuing "usefulness", making image generation a truly deployable productivity tool.
qwen-image-3.0
Clear instruction understanding: supports up to 4.5k token input, generating complex image-text instructions in one pass. Stable text rendering: 10px small text, 12 languages, 20+ fonts clearly legible, ready for infographics and interfaces. Handy batch output: posters, web pages, interfaces and other daily tasks can be generated in bulk at lower cost, ideal for continuous creation. Qwen-Image-3.0 Standard pursues not only generation quality but also everyday creative convenience, making image generation a sustainable content productivity tool.
qwen3.8-max
A 2.4-trillion-parameter MoE model with a comprehensive leap in coding and office capabilities, capable of autonomously coding for days to deliver complete projects. It handles hundreds of professional tasks such as legal, financial, and design work, delivering production-grade results end-to-end in a single conversation. Native visual understanding runs through the full planning, execution, and verification pipeline, supporting deep semantic parsing of ultra-long documents and long videos. It performs autonomous planning and closed-loop iteration in long-horizon tasks, continuously evolving.
deepseek-v4-flash-0731
A highly efficient, lightweight MoE model with 284 billion parameters in total and 13 billion activated parameters, natively supporting context windows of up to one million tokens. It offers fast inference speed, low latency, and cost-effective invocation, delivering well-balanced overall performance. Designed for high-concurrency, lightweight workloads, it is ideally suited for common, essential use cases such as everyday dialogue, content creation, basic RAG applications, and batch text processing.
qwen3.7-flash
qwen3.7-flash-2026-07-15
The Qwen3.7 native vision-language series Flash models comprehensively enhance multimodal understanding and Agent execution capabilities compared to 3.6-Flash. Key improvements include strengthened foundational multimodal abilities, enhanced object recognition, improved real-world perception and spatial intelligence. Multimodal Agent scenarios such as Search Agent and CI Agent have seen significant upgrades, with more stable end-to-end task execution. Multimodal coding capabilities are optimized, delivering a smoother vibe coding experience.
qwen-plus-character
The role-playing model of the Qwen series. This is a dynamically updated version, and notifications will be provided in advance for any model updates. It is suitable for anthropomorphic role-playing and has optimized capabilities in following predefined character instructions, advancing conversations, and demonstrating active listening and empathy. Additionally, it supports the deep restoration of personalized characters.
glm-5.2
GLM-5.2 is a model from Zhipu AI, designed for long-horizon tasks with support for an ultra-long 1M context window. It features powerful logical reasoning, long-text comprehension, and code generation capabilities, balancing performance with inference efficiency. It excels across multi-task benchmarks and is well-suited for intelligent interaction, enterprise applications, and development assistance scenarios.
kimi-k2.7-code
kimi-k2.7-code is Kimi's coding model. It follows instructions more reliably over long contexts and completes programming tasks with higher success rates. It supports text, image, and video inputs, along with thinking mode, conversation, and agent tasks.
qwen3.7-max-2026-06-08
The Max model, the largest in the Qwen3.7 series, has added visual‑modal understanding compared to the May 20 snapshot, enabling it to perceive real‑world scenes and supporting multimodal interactive hybrid agent capabilities. This version is based on a snapshot taken on June 8, 2026.
qwen3.7-plus
qwen3.7-plus-2026-05-26
Among the Qwen3.7 series, the cost-effective Plus model builds on its robust text capabilities while delivering a comprehensive upgrade to its vision‑language abilities, all while preserving its full‑stack agent‑level intelligence for coding, tool use, and productivity workflows. Its key distinguishing feature is multi‑modal interactive hybrid agent capabilities, enabling it to perceive real‑world scenes, read screens and interact with GUIs, generate code based on visual references, and perform end‑to‑end navigation within mobile apps.
qwen3.7-max
qwen3.7-max-2026-05-20
The Max model, the largest in the Qwen3.7 series, currently offers a pure‑text‑only interface for public experimentation. Qwen3.7 is a next‑generation model designed for the agent‑centric era, with its core strengths lying in the breadth and depth of its agent‑level capabilities: it excels at programming, office and productivity tasks, and long‑term autonomous execution.
qwen3.6-flash
The Qwen3.6 native vision-language Flash model series delivers a significant performance boost over the 3.5-Flash version. This model particularly excels in agentic coding capabilities, substantially outperforming its predecessor on multiple code-agent benchmarks, as well as in mathematical and code reasoning. In terms of vision, it features markedly improved spatial intelligence, with especially notable enhancements in object localization and object detection.
qwen3.6-plus
The Qwen3.6 native vision-language Plus series models demonstrate exceptional performance on par with the current state-of-the-art models, with a significant improvement in overall results compared to the 3.5 series. The models have been markedly enhanced in code-related capabilities such as agentic coding, front-end programming, and Vibe coding, as well as in multi-modal general object recognition, OCR, and object localization.
qwen3.5-flash
qwen3.5-flash-2026-02-23
The Qwen3.5 native vision-language Flash models are built on a hybrid architecture that integrates a linear attention mechanism with a sparse mixture-of-experts model, achieving higher inference efficiency. Compared to the 3 series, these models deliver a leap forward in performance for both pure text and multimodal tasks, offering fast response times while balancing inference speed and overall performance.
qwen3-max-2026-01-23
Compared with the snapshot as of September 23, 2025, the Qwen-3 series Max model in this release achieves an effective integration of thinking and non-thinking modes, resulting in a comprehensive and substantial improvement in the model's overall performance. In thinking mode, the model simultaneously supports web search, web information extraction, and a code interpreter tool, enabling it to tackle more complex and challenging problems with greater accuracy by leveraging external tools while engaging in slow, deliberative reasoning. This version is based on a snapshot taken on January 23, 2026.
2025
qwen3-vl-plus-2025-12-19
The Qwen3 series of visual understanding models effectively integrates thinking and non-thinking modes. Compared to the snapshot released on September 23, this version delivers superior performance in reasoning and analysis tasks as well as style control, while also offering lower latency and faster response speeds. This version is based on a snapshot taken on December 19, 2025.
qwen-plus-2025-12-01
This version is a snapshot as of December 1, 2025, and features improved reasoning capabilities compared to the July 28 snapshot. Agent capabilities and multi-turn tool invocation abilities have been further enhanced, and performance on subjective creative tasks has improved significantly. It supports a context length of up to 1 million tokens, with tiered pricing based on context length.
qwen3-max
The Qwen 3 series Max model has undergone specialized upgrades in agent programming and tool invocation compared to the preview version. The officially released model this time has achieved state-of-the-art (SOTA) performance in its field and is better suited to meet the demands of agents operating in more complex scenarios.
qwen3-vl-plus
The Qwen3 series VL models effectively integrates thinking and non-thinking modes, achieving world-leading performance in visual agent capabilities on public benchmark datasets such as OS World. This version features comprehensive upgrades in areas like visual coding, spatial perception, and multimodal reasoning, significantly enhancing visual perception and recognition abilities, and supporting the understanding of ultra-long videos.
qwen-plus
Qwen-Plus is an enhanced version of the Qwen ultra-large language model that supports multiple input languages such as Chinese and English. Compared to previous versions, it shows significant improvements in both Chinese and English code generation, logical reasoning, and multilingual abilities. The response style has been greatly adjusted to align with human preferences, with noticeable enhancements in the level of detail and clarity of responses. Specialized improvements have been made in creative writing, adherence to JSON formatting, and role-playing abilities.
text-embedding-v4
The General Text Vector V4 version is a multi-language text vector model developed by the Tongyi Lab based on Qwen3. Compared to the V3 version, it significantly improves performance in text retrieval, clustering, and classification tasks. It achieves a 15% to 40% improvement in evaluation tasks such as MTEB multilingual, Chinese-English, and code retrieval. Additionally, it supports user-defined vector dimensions ranging from 64 to 2048.