Mistral Models Versioning And Deployment

Production pins the model id to a floating alias (e.g., 'mistral-large-latest'); a silent version bump changes outputs and breaks a downstream parser.

Pin to a dated/versioned model id in production for reproducibility; treat aliases like '-latest' as moving targets and upgrade deliberately behind an eval gate. Roll forward, don't drift.

Operator uses a general chat model for fill-in-the-middle code completion and gets poor results, unaware Codestral and the /v1/fim/completions endpoint exist.

Route code-completion / fill-in-the-middle tasks to Codestral via the dedicated FIM endpoint with prompt+suffix; use general chat models for conversational tasks. Match the model and endpoint to the task.

Code hard-codes a list of available model ids that goes stale as models are added/retired.

Query GET /v1/models for the authoritative list of available models and their ids/capabilities rather than hard-coding; refresh on deploy. Treat any embedded list as a cache, not the source of truth.