DeepSeek V4.1 Flash scheduled for retirement: migrate to version 1017

Databricks announced the retirement of the open model DeepSeek V4.1 Flash, scheduled for November 19, 2026. The recommended replacement is DeepSeek V4.1 Flash (1017).

Use cases

Imagine you run batch inference workloads using ai_query or serve endpoints via Foundation Model APIs. Finding and tracking which pipelines rely on this model lets you update configurations ahead of time, preventing downtime when the model is removed.

How to use

You can audit which endpoints and users are actively calling the retiring model by querying Databricks system tables:

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 '%deepseek-v4.1-flash%'
 GROUP BY eu.requester, se.endpoint_name, se.entity_name
 ORDER BY request_count DESC;

What happens in this code:

1. Joins endpoint consumption metrics (endpoint_usage) with served model metadata (served_entities).
2. Filters queries targeting the specific model name scheduled for retirement.
3. Groups records by requester and endpoint to show request volumes, token counts, and the latest access timestamp.

Limitations and requirements

• Status: Deprecated. Complete retirement is scheduled for November 19, 2026.
• Applies to pay-per-token, provisioned throughput, and ai_query batch inference.
• Workspaces not actively using this model prior to deprecation cannot provision new workloads with it.
• On the retirement date, all remaining workloads pointing to the deprecated model will fail.