Generative AI models maintenance policy

This article describes the model maintenance policy for the Foundation Model APIs pay-per-token and Foundation Model APIs provisioned throughput offerings.

In order to continue supporting the most state-of-the-art models, Databricks might update supported models or retire older models for these offerings.

Model retirement policy

The following sections summarize the retirement policy for the indicated feature offerings.

important

The retirement policies that apply to the Foundation Model APIs pay-per-token and Foundation Model Fine-tuning offerings only impact supported chat and completion models.

Foundation Model APIs pay-per-token

The following table summarizes the retirement policy for Foundation Model APIs pay-per-token.

Retirement notification	Transition to retirement	On the retirement date
Databricks takes the following steps to notify customers about a model that is set for retirement: On the Serving page of your Databricks workspace, a warning message appears on the model card that indicates that the model is planned for retirement. The applicable documentation contains a notice that indicates the model is planned for retirement and the start date it will no longer be supported.	Databricks will retire the model in three months. During this three-month period, customers can either: Choose to migrate to a Foundation Model APIs provisioned throughput endpoint to continue using the model past its end-of-life date. Migrate existing workflows to use recommended replacement models.	The model is no longer available for use and removed from the product. Applicable documentation is updated to recommend using a replacement model.

Foundation Model APIs provisioned throughput

The following table summarizes the retirement policy for Foundation Model APIs provisioned throughput.

Retirement notification	Transition to retirement	On the retirement date
Databricks takes the following steps to notify customers about a model that is set for retirement: For endpoints that serve a deprecated model, a warning message appears on that serving endpoint's details page in your Databricks workspace. This message indicates that the model is planned for retirement and the applicable retirement date. A tooltip message provides recommended alternate models for workload migration. The applicable documentation contains a notice that indicates the model is planned for retirement and the start date it will no longer be supported.	Databricks will retire the model in six months. During this six-month period: Customers can continue running existing provisioned throughput endpoints using the deprecated model until the retirement date. All stopped endpoints that use deprecated models cannot be restarted. Customers that are not actively using a deprecated model cannot create new provisioned throughput endpoints for a deprecated model.	The model is no longer available for use and removed from the product. All endpoints using the retired model are transitioned to a failed state with a descriptive message. Any requests to these endpoints will fail. The customer can delete endpoints that use the retired model, but cannot restart them. Applicable documentation is updated to recommend using a replacement model.

Model updates

Databricks might ship incremental updates to pay-per-token models to deliver optimizations. When a model is updated, the endpoint URL remains the same, but the model ID in the response object changes to reflect the date of the update. For example, if an update is shipped to meta-llama/Meta-Llama-3.3-70B on 3/4/2024, the model name in the response object updates to meta-llama/Meta-Llama-3.3-70B-030424. Databricks maintains a version history of the updates that you can refer to.