Mistral AI Unveils 'Devstral', an Agentic LLM for Software Engineering

Devstral, born from a collaboration between Mistral AI and All Hands AI, is an agentic LLM that performs software engineering tasks. Unlike existing LLMs that excel at atomic coding tasks such as writing a single function, Devstral is designed to solve real development problems, such as understanding the context of large codebases and finding complex bugs.

On the SWE-Bench Verified benchmark, Devstral achieved a score of 46.8%, surpassing existing open-source SOTA models by more than 6 percentage points. In particular, it outperformed larger models such as Deepseek-V3-0324 and Qwen3 232B-A22B, and achieved performance more than 20% higher than GPT-4.1-mini.

Key features include the following:

• Efficient local deployment: It can run on a single RTX 4090 or a Mac with 32GB RAM, making it optimized for on-device and local environments.
• Enterprise suitability: Enables agentic coding even on private repositories at enterprises where security and compliance are critical.
• Versatile usability: Can be integrated with code agent scaffolds such as OpenHands to resolve real GitHub issues.

Devstral is released for free under the Apache 2.0 license and is immediately available via HuggingFace, Ollama, Kaggle, Unsloth, and LM Studio. It is also available through the Mistral API.