Model platform feature updates
For Agent feature updates, see Agent platform updates.
2026
Model upgrade notice
Model upgrade notice Learn more
Alibaba Cloud Model Studio Personal Token Plan user benefits upgrade
Alibaba Cloud Model Studio Personal Token Plan user benefits upgrade Learn more
Memory commercialization notice
Memory commercialization notice Learn more
Legacy Enterprise Knowledge Base sunset notice
Legacy Enterprise Knowledge Base sunset notice Learn more
Managed Agent commercialization notice
Managed Agent commercialization notice Learn more
GLM-5.2 Fast mode price reduction notice
GLM-5.2 Fast mode price reduction notice Learn more
Gateway change notice
Gateway change notice Learn more
Sunset notice for select legacy models
Sunset notice for select legacy models Learn more
Sunset notice for select legacy long-tail models
Sunset notice for select legacy long-tail models Learn more
Delisting notice for Model Studio products: Xiyan GBI and Xiyan GBI Service Package
Delisting notice for Model Studio products: Xiyan GBI and Xiyan GBI Service Package Learn more
Extended sunset schedule for select legacy models
Extended sunset schedule for select legacy models Learn more
New shared usage packs for Team Edition
Flexible Credits usage packs shared across seats, which can offset overage usage, Learn More
First-month offer for new Pro plan subscribers
First-time subscriptions to the Coding Plan Pro plan enjoy the first month at ¥39.90, Learn More
Announcement on API Key encrypted storage and workspace-dedicated inference API domain upgrade
Announcement on API Key encrypted storage and workspace-dedicated inference API domain upgrade Learn more
qwen-turbo resource plan sunset initiation notice
qwen-turbo resource plan sunset initiation notice Learn more
New Codex client access
The Codex terminal AI coding assistant can now access Model Studio, Learn More
PTU long input and caching
Preset throughput deployment now supports long input and prefix caching. View details
New regions and deployment coverage
New regions in the US, Germany, and Japan, along with expanded service deployment coverage, Learn More
Coding Plan web search MCP upgrade
The Coding Plan web search MCP has been upgraded to the Streamable HTTP protocol, with the first 2,000 calls free of charge. View details
Adds leaderboard and comprehensive evaluation
The model evaluation service now includes a leaderboard and comprehensive evaluation, supporting scoring methods such as BLEU_4. Learn More
Availability on the international site
The model import feature is now available on the international site, supporting the import of LoRA fine-tuned models from OSS. Learn More
Model import API launched
The API Reference now includes a new section for model import, covering APIs for import tasks and custom model objects. Learn More
New documentation for temporary API keys
New documentation explains how to generate temporary API keys to avoid exposing permanent API keys in untrusted environments. Learn More
Asynchronous calls
The Responses API now supports asynchronous calls. You can submit long-running tasks by setting background=true and then poll for the results. Learn More
New reinforcement learning training (invitation-only)
Introduces reinforcement learning (RL) training to optimize model strategies using reward signals. This feature is currently invitation-only. Learn more
Support for image generation models
Model tuning supports custom training for image generation models, such as Wan / Wanx. Learn more
Tongyi Multimodal Translation API reference now available
The API reference for Tongyi Multimodal Translation is now available, with APIs for text, image, document, and webpage translation, Learn more
New model compression module
The new model compression module uses quantization algorithms to convert fine-tuned, full-precision models into low-precision versions, reducing deployment costs. Learn more
Text generation API endpoint updated
The text generation API endpoint now includes categories for the OpenAI Responses and Anthropic Messages interfaces. Learn more
Team management for Token Plan Team Edition
The Token Plan Team Edition now includes team management features, supporting SSO/DingTalk logon, seat allocation, and Credits usage monitoring. Learn more
Zero-code safety and compliance enhancement
The model tuning workflow for text generation models now includes a zero-code option to enhance the safety and compliance of large models. Learn more
Added support for EventBridge HTTP callbacks and RocketMQ for asynchronous tasks
Asynchronous tasks can now push completion events through EventBridge, eliminating the need for polling. Learn more
New programming tools, including Kilo CLI
The client development tools now include Kilo CLI, accessible through Token Plan, Coding Plan, and pay-as-you-go. Learn more
New documentation for CosyVoice Voice Cloning API
Detailed documentation for the CosyVoice Voice Cloning API is now available, covering how to create and query voice cloning tasks. Learn more
Deploy new pre-trained models via the API with time-based billing per Model Unit (MU)
The model deployment API now supports deploying pre-trained models such as qwen-flash and qwen-plus, and introduces time-based billing per Model Unit (MU). Learn More
Support for visual understanding (VL) models
Model fine-tuning now supports the visual understanding (VL) model type, enabling you to perform customized training on multimodal models. Learn More
Support for video generation models
Model fine-tuning now supports the video generation model type, enabling you to perform customized training on video models such as the Wan series. Learn More
2025
New dashboard for free tier and usage statistics
Provides a centralized view of each model's free tier, including used/remaining quotas and call volume statistics. model usage
New leaderboard feature
You can now select multiple models at once and generate a comparison leaderboard based on evaluation metrics like BLEU and ROUGE for easy side-by-side performance comparison. model evaluation
New evaluator types
Supports comprehensive model performance assessment across mainstream evaluation scenarios, including string matching, text similarity calculation, model-based and manual classification, and model-based scoring. model evaluation
Added support for viewing model inference logs
You can now view historical conversations between users and models, which helps with prompt optimization and troubleshooting. model monitoring
New model unit deployment method (time-based billing)
Model Unit core advantages of deployment:
Flexible performance adjustment
High service stability
Predictable fixed costs
New:SFT model fine-tuning and deployment for Qwen3-VL models
Qwen3-VL-8B-Instruct and Qwen3-VL-8B-Thinking now support SFT model fine-tuning and deployment of fine-tuned models.
SFT (Supervised Fine-Tuning) supports both full-parameter fine-tuning and efficient fine-tuning (LoRA) methods.
DPO preference training support for Qwen2.5 and Qwen3 series
The Qwen3-32B, 14B, 8B models and the Qwen2.5-72B, 32B, 14B, 8B models now support DPO preference training.
DPO preference training uses negative feedback to reduce hallucination and better align model outputs with human preferences.
Qwen2.5-VL models now support fine-tuning and deployment
The Qwen2.5-VL-72B, 32B, and 7B models now support model fine-tuning and deployment.
New "stop when free tier is depleted" feature
The stop when free tier is depleted feature, when enabled, prevents users from making calls after their free tier is exhausted. It returns an error with the code AllocationQuota.FreeTierOnly and prevents additional charges.
New alerts and notifications
The system notifies you or your operations team when a configured metric, such as call statistics or performance indicators, becomes abnormal. model monitoring
New advanced monitoring mode
Advanced monitoring provides minute-level data refresh with low latency and records failed model call details, including 4xx and 5xx error counts. model monitoring.
Billing adjustment for DeepSeek series models
Some models have transitioned from free trial to paid service. DeepSeek on Alibaba Cloud.
Price reduction for the qwen-max model
The input price for the qwen-max model has been reduced by 88%, and the output price has been reduced by 84%:
For real-time calls, the input price is reduced from CNY 0.02 per 1,000 tokens to CNY 0.0024 per 1,000 tokens, and the output price is reduced from CNY 0.06 per 1,000 tokens to CNY 0.0096 per 1,000 tokens.
For batch calls, the input price is reduced from CNY 0.01 per 1,000 tokens to CNY 0.0012 per 1,000 tokens, and the output price is reduced from CNY 0.03 per 1,000 tokens to CNY 0.0048 per 1,000 tokens.
For the latest billing details, see the Model Studio console.
Price adjustment for some Qwen series models
The prices for the qwen2.5-14b-instruct and qwen2.5-7b-instruct models are reduced by 50%:
The input price for the qwen2.5-14b-instruct model is reduced to CNY 0.001 per 1,000 tokens, and the output price is reduced to CNY 0.003 per 1,000 tokens.
The input price for the qwen2.5-7b-instruct model is reduced to CNY 0.0005 per 1,000 tokens, and the output price is reduced to CNY 0.001 per 1,000 tokens.
The qwen2.5-3b-instruct and qwen2-vl-72b-instruct models have transitioned from a free trial to a paid service:
The input price for the qwen2.5-3b-instruct model is CNY 0.0003 per 1,000 tokens, and the output price is CNY 0.0009 per 1,000 tokens.
The input price for the qwen2-vl-72b-instruct model is CNY 0.016 per 1,000 tokens, and the output price is CNY 0.048 per 1,000 tokens.
For the latest billing details, see the Model Studio console.
New model monitoring capability
Model monitoring tracks model usage and performance, helping you identify issues and optimize performance. model monitoring.
2024
billing
Price reduction for Qwen-VL models
- Qwen VL model prices reduced by up to 85%. Qwen-VL Series Model Price Reduction Notice.
model fine-tuning
SFT fine-tuning for qwen2.5-7b-instruct
- The qwen2.5-7b-instruct model now supports both full-parameter and efficient SFT fine-tuning.
model deployment
Model deployment now supports pay-per-invocation billing
- The pay-per-invocation billing option now supports the deployment of fine-tuned models, including qwen2.5-7B, 14B, 32B, 72B, and qwen2-7B.
ModelScope
New context cache feature
- Context cache reduces redundant inference computations, improving response speed and lowering costs without affecting quality. context cache.
Currently, only the qwen-plus model is supported.
OpenAI API compatibility
Batch jobs now support task notifications.
Two more models are now supported for batch jobs.
- Batch jobs now support completion notifications via callbacks or EventBridge messages, eliminating the need to poll for status. OpenAI Compatibility - Batch (File Input).
- Newly supported models: qwen-vl-max, qwen-vl-plus, and qwq-32b-preview. For billing details, see the Model Studio console.
ModelScope
New search_options parameter
- You can now use the search_options parameter to configure settings for web search, such as search sources and the number of results. This parameter is available for the qwen-max, qwen-plus, and qwen-turbo models. For usage instructions, see the Qwen API Documentation - DashScope.
Data processing now supports canvas orchestration
With canvas orchestration, you can flexibly combine data cleaning nodes and data augmentation nodes to easily create complex data processing workflows.Data cleaning or augmentation
model sunsetting
The qwen-max-longcontext model will be sunset on October 17, 2024. We recommend migrating to qwen-max.
Prepaid support for model inference
Savings plans (prepaid) are now available to offset inference fees beyond the free tier for Qwen, Tongyi Farui, Baichuan-Open Source, ChatGLM, and OpenNLU models. Prepaid (savings plan).
Free tier validity period extended
Free tier validity period for select models extended from 30 to 180 days. notice on the free tier validity period adjustment for Qwen series models.
Price reduction for select Qwen series models
Prices for qwen-max, qwen-turbo, and qwen-plus have been reduced. Qwen series model price reduction notice.
New model types added
New models have been added, including Qwen Coder, Qwen Audio, and the open-source version of Qwen 2.5. You can find these new models in the model list.
Support for custom quota increase requests
Submit a custom quota increase request from a model's details page in Model Square or directly from the quota increase request page.
Model Studio console redesign
The console has been redesigned for a clearer and more user-friendly interface. Explore the new capabilities in the model playground.
Model training billing adjustment
Model training billing updated: hybrid training is now a billable item. Formula: (Total training tokens + Total hybrid training tokens) × epoch × Unit price. billable item.
New models supported for model training
qwen-turbo-0624 and qwen-plus-0723 now support model training. For billing rules, see the Model Studio console.
Billing adjustment for large models
The qwen-vl-max-0809, qwen-vl-max-0201, sensevoice-v1, and paraformer-realtime-v2 models are now billable. For billing rules, see the Model Studio console.
Model data
Enhanced model data import
- Improved error details for failed and partially successful imports.
- Status polling is now available for model data publishing and imports.
OpenAI API compatibility
New Vision invocation mode
- The new Vision invocation mode lets you use the Qwen vision model by simply adjusting the API-KEY, BASE_URL, and model parameters in your existing framework. This enables seamless integration with OpenAI interfaces and tools. To get started, see OpenAI Compatibility - Vision.
OpenAI API compatibility
New batch invocation mode
- The new batch invocation mode lets you send model requests in batches and receive results asynchronously within 24 hours at half the cost of real-time invocations. To get started, see OpenAI Compatibility - Batch (File Input). For billing information, see the Model Studio console.
This mode is ideal for high-volume, asynchronous inference tasks, such as data processing and cleansing with large models, large-scale data analysis, and automated extraction.
Model tools
New model evaluation method
- Automated model evaluation is now available. Introduction to model evaluation.
Model tools
New models for dedicated deployment
- The qwen2-7b-instruct model now supports custom training and dedicated deployment.
- The qwen2-72b-instruct model now supports custom training and dedicated deployment. To get started, see Fine-tune models in the console.
New operator types for data processing
Added the detoxification operator, which automatically detects, analyzes, and removes sensitive or non-compliant content from your data. data cleaning or augmentation
New operator types for data processing
Added the sensitive word filtering operator, which filters out samples containing sensitive words. data cleaning or augmentation
Updated query capability for call statistics
Sub-workspaces can now view only the details of their own business data. model monitoring
Added best practice documentation
The console now includes best practice code samples.
Upgraded Qwen series models
Improved the model performance of qwen-plus&qwen-turbo. Click text generation to view details.
Added support for model fine-tuning for qwen-vl-plus
qwen-vl-plus now supports model fine-tuning using multimodal training sets. training set and evaluation set.
Data augmentation enhancements
Data augmentation now supports parameter and prompt configuration for more effective data processing.
data management supports data filtering
data management now supports batch deletion and data filtering by file name or file status.
Price reduction for select models and separate billing for input and output
Input and output are now billed separately for select Qwen models, with an across-the-board price reduction. For details, see the Tongyi Qwen series large model price adjustment notice.
Free tier policy update for new users
New users receive a total of 36 million free tokens for mainstream Tongyi Qwen large models. (This promotion is no longer available.)
A redesigned Home page that improves the onboarding experience
You can make up to 100 API calls to the max, turbo, and plus models before activating the service.
The homepage has been streamlined for quick onboarding. Click QuickStart to view details.
A completely redesigned, more intuitive, and clearer interface
The qwen-max model has been fully upgraded.
The redesigned interface supports filtering models by series and provides a global search filter.
View API call examples with a single click to quickly invoke models.
A one-stop workflow now covers the entire model lifecycle: invocation, model fine-tuning, billing, authorization, and service activation.
New users can try out select models directly online
Try models before activating the service.
Optimized the model selection mechanism to support selection by series and custom models.
View API call examples with one click.
The VL and Wan models are now fully available for online testing. Click Introduction to Model Playground for details.
New model types available for training
SFT and deployment are now available for the qwen-plus and qwen1.5-72B models.
SFT and deployment are now available for the hundred-billion-parameter open-source model, qwen1.5-110B. Fine-tune Models in the Console.
New automated data processing capabilities and enhanced data security
Automated data processing now includes operators for data deduplication, sensitive information removal, and more.
One-click processing for SFT data is now available to improve data quality and address issues such as privacy leaks and data duplication in training sets.
You can now create datasets of different types, such as training and evaluation sets, and manage multiple data versions. Published datasets can be used for model fine-tuning or evaluation.
New multi-version management for model data enables a more efficient workflow.
Optimized the data display and configuration logic.
Service agreement update
The updated service agreement will take effect on May 2, 2024. View the full Alibaba Cloud Model Studio Service Agreement.
New models available for testing
Qwen VL models are now available for interaction in the Playground. Click Qwen to view details.
The Wanx model is now available for testing in the model experience center. Click Tongyi Wanxiang for details.
Improved the auto-scroll-to-bottom behavior in the model experience center.
Model deprecation and upgrade
The qwen-max-1201 model will be deprecated on April 22, 2024. We recommend migrating to qwen-max. Qianwen model version upgrade instructions.
Module updates for Alibaba Cloud Model Studio
Added Qwen series & third-party open-source large language models.
Added model hub content updates. Click model to view details.
Updated the App Center. See Application Development for details.
Updated System Management. See System Management for details.
New large models and billing adjustments
Added Qwen series large models.
Adjusted large model billing. Notice on Billing Adjustments for Qwen Series Large Models.
Added the model experience center
Supports multi-model comparison and model debugging. For details, see Model Experience Introduction.
Added large model types
Added official and third-party large language models. For details, see Models.
Added the Qwen-Plus model deployment type
Supports independent deployment for the Qwen-Plus large model.
New option for mixed training
Mixed training is now supported. You can specify a data ratio to automatically mix data for training.
New offline status
You can now delete a dedicated model instance that is offline.
New models supported for training
You can now train the Qwen-Plus and Qwen-Turbo models.