Devstral Medium

Devstral Medium sits within Mistral's devstral family and is positioned as a high-performance agentic coding model. According to Puter Developer's catalog entry, the model achieves 61.6% on SWE-Bench Verified, a benchmark that measures a model's ability to resolve real-world GitHub issues end-to-end, signaling a focus on software engineering tasks rather than general chat. Practical integration guides route it through Mistral's chat completions endpoint with a developer-named identifier that follows the "devstral-medium-2507" naming convention, treating it as a text-in, text-out model designed for tool-use workflows in agent environments.

For teams building coding assistants, Devstral Medium offers a long context window suited to ingesting sizeable codebases, paired with tool-calling and temperature-control capabilities that let agents invoke external functions and tune response randomness. The combined context-plus-output ceiling gives enough room to hold multi-file repositories alongside generated patches in a single turn. Its pricing tier and modality profile position it as a workhorse for production agent pipelines, particularly where benchmark-driven SWE performance matters more than open-weight self-hosting, since third-party listings describe it as a hosted Mistral offering rather than a downloadable release.