Foundation models maintenance policy

This article describes the model maintenance policy for the Foundation Model APIs pay-per-token and Mosaic AI Model Training offerings.

In order to continue supporting the most state-of-the-art models, Databricks might retire older models or update supported models for the Foundation Model APIs pay-per-token and Mosaic AI Model Training offerings.

Model retirement policy

The following retirement policy only applies to chat and completion models.

If a model is set for retirement, Databricks takes the following steps to notify customers:

  • A warning message displays in the model card from the Serving page of your Databricks workspace that indicates that the model is planned for retirement.

  • A warning message displays in the dropdown menu for Mosaic AI Model Training in the Experiments tab that indicates that the model is planned for retirement.

  • Applicable documentation contains a notice that indicates the model is planned for retirement and the start date the model will no longer be supported.

After customers are notified about the upcoming model retirement, Databricks will retire the model in 3 months. During this period of time, customers can choose to migrate to a provisioned throughput endpoint to continue using the model past its end-of-life date.

See Retired models for a list of currently retired models and planned retirement dates.

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 llama-2-70b-chat on 3/4/2024, the model name in the response object updates accordingly to llama-2-70b-chat-030424. Databricks maintains a version history of the updates that customers can refer to.

Retired models

The following table shows recently announced model retirements along with the retirement dates and recommended replacement models to use for Foundation Model APIs pay-per-token workloads. Databricks recommends that you migrate your applications to use replacement models before the indicated retirement date.

Model

Retirement date

Recommended replacement model

MPT 7B Instruct

August 30, 2024

Mixtral-8x7B

MPT 30B Instruct

August 30, 2024

Mixtral-8x7B

If you require long-term support for a specific model version, Databricks recommends using Foundation Model APIs provisioned throughput for your serving workloads.