codestral-2508

Overview. Codestral 2508 is Mistral AI's closed-weight, API-only code model, purpose-built for fill-in-the-middle completion, code correction, and test generation. Released in August 2025, it ships with a 256K-token context window, making it a candidate for teams that need a coding-focused model to reason across large files or multi-file projects rather than short snippets. It is served exclusively through Mistral's hosted API — there are no public weights.

Key strengths. The model is tuned for the editor loop rather than for open-ended chat: it accepts text and code files as plain-text input and returns text, and it exposes tool calling and structured outputs for schema-constrained decoding. On Design Arena its Elo ranges from 1034 in game development to 1099 in 3D, with UI components at 1072. It is not a reasoning model — configurable reasoning effort is not offered here, and Mistral reserves that control for its separate hybrid and reasoning lines.

When to use it. Reach for Codestral 2508 for fill-in-the-middle autocomplete, targeted code correction, and automated test generation, especially where the task spans a large codebase and needs the full 256K context rather than a single file. Its interactive latency profile and tool-calling support also make it a reasonable fit for chat-assisted coding sessions that call external tools or expect structured, JSON-schema-validated responses. It is not positioned as a general-purpose reasoning or extended-thinking model, so keep it scoped to code-centric tasks.

Limitations. Mistral has not disclosed parameter count, architecture, or licence for this version across repeated verification passes, so those remain unknown to customers. It has no vision input — the 'file' input type covers code, text, and PDFs read as plain text, not image understanding — and its Design Arena code-quality signal is mid-tier and has been declining over recent cycles, so it should not be treated as a frontier coding model.

Sovereignty note. Codestral 2508 is built by Mistral AI, a French company, and is available only as a hosted API, with no public weights. Where requests run, and whether inputs are retained or used for training, depends on the endpoint serving it; see the endpoint facts on this page.