Codestral and Devstral

Codestral and Devstral are coding models for developers who want to run completion tools or software engineering agents on their own hardware. The associated mistral-inference library is archived and no longer maintained.

• Code completion: Codestral supports fill-in-the-middle completion to fill gaps between existing sections of code.
• Repository edits: Devstral powers agents that use tools to inspect codebases and make changes across multiple files.
• Image input: Devstral Small 2 accepts images as well as text, letting coding assistants analyze visual material.
• Agent compatibility: Devstral works with Mistral Vibe, Cline, Kilo Code, Claude Code, OpenHands and SWE Agent.

Downloaded weights let you run inference locally or on your own server. Mistral also offers a hosted API, where inference runs on its cloud infrastructure. Devstral supports vLLM, SGLang, Transformers, llama.cpp and Ollama; Devstral Small 2 also works with LM Studio. Community GGUF weights are available for llama.cpp. Devstral Small 2 can run on a single NVIDIA RTX 4090 or a Mac with 32GB RAM.

Licenses differ by model. Devstral Small 2 uses Apache 2.0, allowing commercial use and modification. Codestral 22B uses the Mistral AI Non-Production License, which restricts production use. The archived inference library has an Apache 2.0 license separate from the licenses of the model weights.