Devstral

Mistral AI teamed up with All Hands AI to build Devstral, a model aimed squarely at the messy, unglamorous work of real software engineering rather than the party trick of writing a clean standalone function on demand. That distinction matters more than it sounds. Most LLMs can spit out a sorting algorithm or autocomplete a loop just fine, but ask them to trace a bug through a sprawling codebase with dozens of interdependent files, and they tend to fall apart.

Devstral was trained specifically to chew through actual GitHub issues, running inside agent scaffolds like OpenHands or SWE-Agent that let it read, test, and iterate against real repositories. On SWE-Bench Verified, a set of 500 hand-checked GitHub problems, it scored 46.8%, beating the previous best open model by more than six points. That's notable given who it's beating: under the same OpenHands scaffold, Devstral outperformed Deepseek-V3-0324, a 671-billion-parameter model, and Qwen3's 232B-A22B variant, both of which dwarf it in size. Against closed models evaluated on their own custom setups, Devstral still came out ahead of GPT-4.1-mini by more than 20 points, which is the kind of gap that makes a small, open model genuinely awkward for the commercial labs to explain away.

Size is really the punchline here. Devstral runs on a single RTX 4090 or a 32GB-RAM Mac, which pushes it into a category most coding agents can't touch: fully local, on-device software engineering. That's a real draw for companies with strict compliance rules who can't ship proprietary code to a third-party API, and it's just as appealing for developers who want an agentic coding assistant living entirely on their own machine, tied into tools like OpenHands.

Mistral is releasing Devstral under Apache 2.0, no strings, so anyone can fine-tune or fold it into their own tooling. It's also live on the API as devstral-small-2505, priced identically to Mistral Small 3.1 at $0.10 per million input tokens and $0.30 per million output tokens, and downloadable through Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio. Mistral is calling this a research preview and says a bigger agentic coding model is already in the pipeline for the coming weeks, which suggests this release is as much a marker in the ground as it is a finished product.