Devstral

Devstral targets agentic coding: Mistral AI built it with All Hands AI, the team behind OpenHands, to explore codebases, edit multiple files, and drive software engineering agent scaffolds rather than autocomplete snippets. Devstral Small, a 24B-parameter model with a 128K context window, scores 53.6 percent on SWE-Bench Verified, which made it the top open-source model on that benchmark at release, ahead of much larger models such as DeepSeek-V3-0324 under the same scaffold, and it is light enough to run on a single RTX 4090 or a Mac with 32 GB of RAM. Weights are Apache-2.0 on Hugging Face, so commercial use and fine-tuning are unrestricted, and deployment is supported through vLLM (recommended), Ollama, llama.cpp, LM Studio, mistral-inference, and Transformers, with a published system prompt file tuned for agent-based interactions. Paired with OpenHands or similar agent frameworks it handles automated issue resolution, multi-file edits, and repository-scale code exploration, making it one of the most practical models to run locally for software engineering agents rather than chat-style code assistance.