Mistral AI Unveils Devstral for Software Engineering

Mistral AI has introduced Devstral, a new agentic large language model (LLM) designed for software engineering tasks. Developed in collaboration with All Hands AI, Devstral outperforms all open-source models on the SWE-Bench Verified benchmark by a significant margin. Released under the Apache 2.0 license, Devstral is available for free to the community.

Devstral is engineered to address real-world software engineering challenges, such as contextualizing code within large codebases and identifying subtle bugs. It achieves a score of 46.8% on the SWE-Bench Verified benchmark, surpassing previous state-of-the-art models by over 6 percentage points. Notably, Devstral outperforms larger models like Deepseek-V3-0324 and Qwen3 232B-A22B when evaluated under the same test scaffold.

The model is lightweight enough for local deployment on devices like a single RTX 4090 or a Mac with 32GB RAM, making it suitable for both individual and enterprise use. Devstral is available for download on platforms such as HuggingFace and can be accessed via Mistral AI's API under the name devstral-small-2505.

Mistral AI encourages feedback on this research preview as they continue to develop larger agentic coding models.