Mistral AI launches Devstral, a powerful open source LLM for coding and engineering tasks

French artificial intelligence startup Mistral has introduced Devstral, a specialized agentic large language model (LLM) designed for advanced software engineering tasks. Developed in collaboration with All Hands AI, the maker of OpenHands, Devstral outperforms all major open-source models on the SWE-Bench Verified benchmark, with results surpassing those of DeepSeek-v3 and Qwen 3. Notably, Devstral also outpaces proprietary offerings such as GPT-4.1-mini and Claude 3.5 Haiku when evaluated on the same metric.

Devstral is engineered to address the challenges of context-rich, real-world development. Unlike typical LLMs that excel primarily at atomic coding tasks, Devstral is tailored for resolving real GitHub issues, contextualizing code across large codebases, and connecting disparate software components. The model operates effectively with code agent scaffolds such as OpenHands and SWE-Agent, which mediate between the model and test cases.

For deployment, Devstral is light enough to run on a single NVIDIA RTX 4090 graphics card or a Mac equipped with 32 GB RAM, making it suitable for local and on-device use. Released under an Apache 2.0 license, Devstral is free for the community to customize and build upon. Distribution is available through Hugging Face, Ollama, Kaggle, Unsloth, LM Studio, and the Mistral API.