Model deprecation policy

Notification process

For snapshot models, which are identified by a specific date in their name (for example, qwen-max-2025-01-25, common for Qwen series models), we issue a sunset notice 30 days before the official sunset date.
For mainline models, which are the core versions of a model series, we issue a sunset notice 3 months before the official sunset date.

We send notifications via email, internal messages, and official website announcements.

Email and internal messages are sent only to users who have called the models scheduled for sunset in the last 3 months.

Starting from the date of the retirement notice, the QPM (queries per minute) and TPM (tokens per minute) for retiring models will gradually decrease. For models that received a quota increase, their limits will revert to the default rate limit before this reduction begins. Throughout this period, the model API and related console features will remain fully functional.
Starting from the official retirement date:
- Model inference: Model inference will be discontinued. API calls to the retired model will fail.
- Model fine-tuning and model deployment: You will no longer be able to start new fine-tuning and deployment operations on the retired model. (For some models, these features may remain available after the retirement date. Please refer to the official retirement notice for details.) This does not affect existing trained and deployed models.
- Console features and official documentation: Associated console features (such as Model Square and Model Discovery) and the official documentation will also be retired.

Check if your account is using any models scheduled for sunset.
If you use an affected model, test the replacement model's business performance before switching to it.

Each announcement is published separately. Click the corresponding link for details.