Mistral AI recently officially launched Devstral, an open-source agent large language model built specifically for software engineering tasks. Developed in partnership with All Hands AI, the model aims to help developers tackle complex, real-world programming problems, supporting tool use and adapting to long-context scenarios.
Model Design Goals
Devstral focuses on agentic coding tasks, capable of understanding context, identifying component relationships, and detecting subtle bugs within large codebases. It is specifically trained to resolve real GitHub issues and runs on code agent frameworks like OpenHands or SWE-Agent to enable end-to-end software engineering workflows.
Exceptional Performance
According to official benchmarks, Devstral achieves a 46.8% resolution rate on the SWE-Bench Verified dataset (containing 500 real GitHub issues), which is over 6 percentage points higher than the best previous open-source model. Within the OpenHands framework, the model surpasses larger models like Deepseek-V3-0324 (671B) and Qwen3 232B-A22B. In several comparisons, it also significantly outperforms GPT-4.1-mini by over 20%.
Open-Source License and Easy Deployment
Devstral is released under the Apache 2.0 license, making it fully open-source and available for commercial use. Developers can download it for free from platforms like Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio. The model is moderately sized and can run locally on a single NVIDIA RTX 4090 GPU or a Mac with 32GB of RAM, meeting enterprise needs in privacy-sensitive scenarios. Additionally, it is available through the Mistral AI API under the name devstral-small-2505, priced the same as Mistral Small 3.1.
Future Outlook
Devstral is currently released as a research preview, and Mistral AI welcomes feedback from the developer community. The company also announced plans to launch a larger-scale agent coding model in the coming weeks, further advancing the application of open-source AI in the field of software engineering.
