MaxCompute models-阿里云帮助中心

Models are a core component of MaxCompute, supporting various types such as public, imported, and remote models. MaxCompute provides unified management and versioning for models, making it easier to integrate AI capabilities into your business analytics workflows.

Model overview

Key concepts

Model: A deployment object registered in MaxCompute for prediction or generation tasks. It seamlessly integrates AI computing capabilities, such as those from large language models (LLMs) and machine learning models, into the same platform where your data resides.
Model version: An independent and uniquely identifiable sub-object of a model. You can create and manage multiple iterative versions under a single model name, which simplifies grayscale releases, rapid rollbacks, and performance comparisons between versions.

Benefits

Unified management: MaxCompute provides multiple model types. Like data, models support permission controls and version control to meet enterprise security and compliance requirements.
Multi-engine support: MaxCompute supports model invocation from various ecosystems, including SQL and Python (MaxFrame). This unified architecture allows data analysts to use familiar SQL to call powerful AI models and enables data scientists to leverage the distributed Python computing power of MaxFrame to improve the efficiency and quality of data preprocessing.
Simplified operations: Perform AI inference directly on your data without exporting it to external systems. This approach avoids the security risks, costs, and latency of data movement.

Model types

MaxCompute provides the following model types:

Type	Description	Tutorial
Public model	MaxCompute provides several built-in, ready-to-use open-source large models. These are pre-created in a public project named `BIGDATA_PUBLIC_MODELSET` and a public schema named default. You can call these models using an AI Function without creating or managing model objects, simplifying access to AI. When you use a public model: If the schema syntax setting is not enabled for the project, use `bigdata_public_modelset.<model_name>`. If the schema syntax setting is enabled for the project, you must specify `bigdata_public_modelset.default.<model_name>`.	Use a MaxCompute public model for sentiment analysis of online reviews
Remote model	Connect to models deployed on PAI-EAS by registering them as MaxCompute remote models. To do this, specify the PAI-EAS endpoint and access token. You can then call the model using an AI Function.	Use a MaxCompute remote model to automatically generate e-commerce product descriptions
Internally trained model	You can use MaxCompute MaxFrame to train traditional machine learning models and run `TO_ODPS_MODEL` to save them as MaxCompute internally trained models.	Use MaxCompute for XGBoost model training and prediction
Imported model	If built-in public models do not meet your specific business needs, you can import custom models. Specify the OSS (Object Storage Service) path to your externally trained and fine-tuned model files to register them for AI inference in MaxCompute.	This feature is being rolled out.

Use the MaxCompute AI Function to call built-in public models or custom models that you manage in your project.

Public models and applicable quotas

Model name	Quota
qwen3.7-max qwen3.7-plus qwen3-vl-embedding text-embedding-v4 qwen3.6-plus qwen3.6-flash deepseek-v4-pro deepseek-v4-flash qwen3.5-397b-a17b qwen3-asr-flash qwen3-max (to be deprecated)	Pay-as-you-go model computing service (Token) Note For more information about model capabilities and billing, see Model computing fees (Token fees).
Qwen3-0.6B-GGUF Qwen3-1.7B-GGUF Qwen3-4B-GGUF Qwen3-8B-GGUF Qwen3-14B-GGUF DeepSeek-R1-Distill-Qwen-1.5B DeepSeek-R1-Distill-Qwen-7B DeepSeek-R1-Distill-Qwen-14B DeepSeek-R1-0528-Qwen3-8B	Pay-as-you-go standard computing resources (CU) Subscription standard computing resources (CU) Note In a production environment, running an AI Function with a CU-based quota can cause performance bottlenecks that affect the throughput and latency of model inference. We recommend activating the pay-as-you-go model computing service (Token) for better performance and flexible scalability.
Qwen3-VL-8B-Instruct	Subscription AI computing resources (GU)

Model management

Before you manage models, make sure that your account has the permissions to manage model objects.

You can manage models in the following ways:

Method	Description
Use SQL to create and manage models	Use SQL statements to manage models, including creating, viewing, modifying, and deleting them.
Use MaxFrame to manage models	The MaxFrame Python library currently supports only model creation.
Use the console to manage models	The MaxCompute console provides a graphical user interface (GUI) for model management. In supported regions, you can use the console to view the models you have created. Procedure: Log in to the MaxCompute console and select a region in the upper-left corner. In the left-side navigation pane, choose Manage Configurations > Projects. On the Projects page, click Manage in the Actions column for the target project. On the Project Settings page, click the Models tab. You can view public models and their version information under the `BIGDATA_PUBLIC_MODELSET` public project. You can also view the custom models you have created and their versions. Note Console-based model management is currently available only in the China (Beijing), China (Hangzhou), China (Shanghai), and China (Shenzhen) regions. This feature is being rolled out to other regions.