Mistral has announced Devstral, a new open-source large language model (LLM) built specifically for code-related tasks.
Developed in collaboration with All Hands AI, the model is now publicly available under the Apache 2.0 license, making it freely usable even for commercial applications.
According to internal benchmarks released by Mistral, Devstral scored 46.8% on SWE-Bench Verified, outperforming other open models such as Google’s Gemma 3 and DeepSeek V3.
The SWE-Bench Verified benchmark tests a model’s ability to resolve real-world GitHub issues, offering a practical measure of code reasoning and problem-solving ability.
Devstral has 24 billion parameters and was trained with a focus on handling complex codebases, navigating multiple files, and working with automated programming agents such as SWE-Agent and OpenHands.
Despite its capabilities, the model is lightweight enough to run locally on an RTX 4090 GPU or even on a Mac with 32GB RAM, making it viable for both enterprise environments and individual developers.
Notably, Devstral outperformed GPT-4.1 mini by a margin of over 20 percentage points on the SWE-Bench Verified benchmark—an unexpected result given its smaller size and open-source nature.
Devstral’s design focuses on real-world development tasks. It supports editing across multiple files and can operate alongside agents like SWE-Agent and OpenHands, aligning with the workflows of modern software teams.
Mistral is treating this version as a research preview, with a larger version already in the works.
In the meantime, Devstral is available via Hugging Face, Ollama, Kaggle, and the company’s API—offered at the same pricing as its earlier Mistral Small 3.1 model.
This release follows a busy few weeks for the Paris-based company, which also launched Mistral Medium 3 and Le Chat Enterprise, a business assistant designed for organizational use.
Founded in 2023, Mistral has already raised over $1.2 billion in funding and counts BNP Paribas, AXA, and other major firms among its clients.
