模型 API
A Batch of Older Mistral Models Will Be Retired on July 31, Requiring Regression Testing for Coding Agents and Structured Outputs
Mistral’s summer model cleanup is entering its final batch, affecting older identifiers for Devstral, Magistral, Mistral Small, and Open Mistral Nemo. The officially recommended replacements are not snapshots of the same architectures, so migration requires more than renaming models: teams must revalidate tool calling, output schemas, context windows, and costs.
Mistral is scheduled to remove a batch of older models on July 31, including `devstral-2512`, `magistral-medium-2509`, `mistral-small-2506`, `magistral-small-2509`, and `open-mistral-nemo-2407`. The recommended replacements span Mistral Medium 3.5, Mistral Small 4, and Ministral 3 8B. This means the change is not a behavior-preserving server-side upgrade, but a remapping of workloads with different use cases, capacities, and price tiers to new models.
The technical risks are particularly significant for coding agents. Devstral was designed for exploring codebases, making cross-file changes, and using tools. After switching to a general-purpose Medium model, function arguments, tool-selection frequency, stopping conditions, and long-trajectory stability may still change even if the API accepts the same message format. Magistral reasoning workloads also cannot be evaluated solely by comparing accuracy on individual questions; teams must test reasoning-token usage, tail latency, and structured-output compliance. Nemo users moving to an 8B-class Ministral model should revalidate language coverage, context limits, and quantized-deployment requirements.
In practice, teams should pin explicit version IDs, replay production traffic to establish differences between the old and new models, and then update routers, cache keys, cost alerts, and fallbacks. If provider endpoints have already stopped responding, teams should also check whether failures are being amplified by incorrect retry behavior. It will be important to monitor whether Mistral provides more complete post-retirement error codes, compatibility windows, and per-model migration benchmarks. Current public information primarily lists replacements and does not demonstrate that the new models are equivalent within existing agent workflows.