Today, Mistral AI and All Hands AI are proud to introduce Devstral, an advanced agentic language model (LLM) specifically designed for software engineering tasks. Devstral stands out by significantly outperforming all existing open-source models on the SWE-Bench Verified benchmark, achieving a score of 46.8%, which is more than 6 percentage points higher than previous state-of-the-art models. It even surpasses much larger closed-source models like Deepseek-V3-0324 (671 billion parameters) and Qwen3 232B-A22B. Addressing Real-World Software Engineering Challenges Traditional LLMs excel at simpler coding tasks such as writing standalone functions or completing lines of code. However, they often fall short when tackling complex, real-world software engineering problems. These challenges include contextualizing code within extensive codebases, identifying relationships between different components, and detecting elusive bugs in intricate functions. Devstral is engineered to overcome these limitations. It is trained on real GitHub issues and integrates with code agent frameworks like OpenHands or SWE-Agent, which define the interaction between the model and the test cases. This training approach ensures that Devstral can effectively handle the nuances of real-world software development. When evaluated using the SWE-Bench Verified benchmark—a dataset comprising 500 manually vetted real-world GitHub issues—Devstral's performance is notably superior. Under the OpenHands test framework, it outperforms not only its open-source counterparts but also several closed-source models, including the recently launched GPT-4.1-mini, by a significant margin. Versatile Deployment Options Devstral's lightweight architecture allows it to run efficiently on a single RTX 4090 GPU or a Mac with 32GB RAM. This makes it an excellent choice for local deployment and on-device use, where it can interact with local codebases and quickly resolve issues. Users can explore the model through our documentation or tutorial videos to get started. In enterprise settings, Devstral is particularly valuable for working with privacy-sensitive repositories. Its robust performance and compliance-friendly design make it suitable for organizations with strict security and regulatory requirements. Developers looking to integrate agentic coding features into their IDEs, plugins, or environments will find Devstral a compelling addition to their model selection. Its versatility and effectiveness in these contexts highlight its potential for enhancing developer productivity. Availability and Accessibility Devstral is freely available under the Apache 2.0 license, empowering the community to build upon, customize, and advance autonomous software development. To get started, refer to our detailed model card. For those preferring to use an API, Devstral is accessible via our API service under the name devstral-small-2505. Pricing matches that of Mistral Small 3.1: $0.1 per million input tokens and $0.3 per million output tokens. If you opt for self-deployment, Devstral can be downloaded from various sources starting today, including HuggingFace, Ollama, Kaggle, Unsloth, and LM Studio. For enterprises needing more tailored solutions, such as fine-tuning on private codebases or integrating Devstral’s capabilities into other models, please reach out to connect with our applied AI team. Future Developments This initial release of Devstral is a research preview, and we are eager to receive feedback from users. Our team is already working on a larger agentic coding model, which we plan to launch in the coming weeks. We are committed to continuously improving our models and exploring new ways to support the software development community. If you are interested in discussing how Devstral can benefit your team or learning more about our range of models, products, and solutions, please contact us. We are excited to collaborate and help drive innovation in the field of autonomous software development.