Generative AI models maintenance policy

This page describes the model maintenance policy for the Foundation Model APIs pay-per-token, Foundation Model APIs provisioned throughput, Batch Inference with ai_query, and Foundation Model Fine-tuning offerings.

To continue supporting the most state-of-the-art models, Databricks manages models through a lifecycle that progresses from update to deprecation to eventual retirement.

• Update: Databricks applies incremental updates to a model to deliver optimizations. See Model updates.
• Deprecated: A deprecated model is no longer recommended for new workloads, but remains available in workspaces with existing usage of the model. Workspaces that aren't using the model at the time of deprecation no longer have access to it.
• Retired: A retired model is no longer accessible, and support for the model is fully discontinued. Any workload using the model stops working.

Model deprecation policy

When Databricks deprecates a model, that model is no longer recommended and is planned for retirement. Databricks announces retirement dates for deprecated models with the notification timelines summarized in the following sections. Retirement dates might be announced at the time of deprecation or at a later date. After the retirement date, the model is no longer accessible, and any workload using it stops working.

For deprecated and retired models and their announced retirement dates, see Deprecated and retired models. For partner models, see Partner model retirement policy.

Foundation Model APIs

The following table summarizes the deprecation policy for the Foundation Model APIs pay-per-token, Foundation Model APIs provisioned throughput, and Batch Inference with ai_query offerings.

Partner model retirement policy

Partner models are models that third-party partners — specifically OpenAI, Anthropic, and Google — provide through Foundation Model APIs. For these partner models, Databricks generally follows the same deprecation timelines and policies described above.

However, partners might provide retirement dates shorter than the three-month transition period that Databricks publishes. In these cases, Databricks attempts to bridge the gap by temporarily redirecting models to a similar version, so customers receive the full transition time.

For example, if a partner model retirement is announced with one month's lead time instead of three, Databricks redirects the model for an additional two months to prevent immediate breakage and allow time for migration. Queries fail at the end of the full three-month period.

Foundation Model Fine-tuning

The following table summarizes the deprecation policy for Foundation Model Fine-tuning.

Model updates

Databricks might ship incremental model updates to deliver optimizations. When Databricks updates a model, 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 Databricks ships an update 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. Reach out to your Databricks account team for more details.

Deprecated and retired models

The following sections list models that are deprecated (no longer recommended for new workloads) or retired (end-of-life and no longer available). Retirement dates for deprecated models are announced at least three months in advance.

Foundation Model APIs retirements

The following table shows model retirements, their retirement dates, and recommended replacement models to use for Foundation Model APIs pay-per-token and provisioned throughput serving workloads. Databricks recommends that you migrate your applications to use replacement models before the indicated retirement date.

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

Foundation Model Fine-tuning retirements

The following table shows retired model families, their retirement dates, and recommended replacement model families to use for Foundation Model Fine-tuning workloads. Databricks recommends that you migrate your applications to use replacement models before the indicated retirement date.

Find workloads that use retired models

Use the following query to find workloads that are using deprecated models and identify their owners.

SELECT
   eu.requester,
   se.endpoint_name,
   se.entity_name,
   COUNT(*) AS request_count,
   SUM(eu.input_token_count) AS total_input_tokens,
   SUM(eu.output_token_count) AS total_output_tokens,
   MIN(eu.request_time) AS first_request,
   MAX(eu.request_time) AS last_request
 FROM system.serving.endpoint_usage eu
 JOIN system.serving.served_entities se
   ON eu.served_entity_id = se.served_entity_id
 WHERE LOWER(se.entity_name) LIKE '%<retired-model-name>%'
 GROUP BY eu.requester, se.endpoint_name, se.entity_name
 ORDER BY request_count DESC