What is Devstral?
For software engineers, tackling complex issues within large codebases often requires more than just generating code snippets. You need an AI that understands context, identifies relationships between components, and helps pinpoint subtle bugs – essentially, an AI that can act as a true engineering partner. Traditional LLMs excel at atomic tasks but often fall short on these real-world challenges.
That's why we're introducing Devstral, an agentic LLM specifically designed for software engineering tasks. Developed in collaboration between Mistral AI and All Hands AI, Devstral was trained on real GitHub issues to handle the intricacies of actual development workflows. It's built to integrate with code agent scaffolds, allowing it to interact with your codebase and help resolve issues effectively.
Devstral is engineered to deliver strong performance where it matters most for developers, offering a powerful tool for automating complex coding tasks and augmenting your engineering capabilities.
Key Features
🤖 Performs Agentic Coding: Trained on real-world GitHub issues, Devstral is built to handle complex, multi-step software engineering tasks beyond simple code generation, working effectively with agent scaffolds.
📊 Delivers Strong Benchmark Performance: Achieves a 46.8% score on SWE-Bench Verified, outperforming prior open-source State-of-the-Art models by over 6% points and surpassing models like GPT-4.1-mini by more than 20% on this challenging dataset.
💡 Lightweight and Accessible: At 24 billion parameters, Devstral is designed for accessibility, capable of running locally on a single RTX 4090 or a Mac with 32GB RAM.
📜 Available under Apache 2.0 License: Offered with a permissive open license, allowing you to use, modify, and build upon Devstral for both commercial and non-commercial purposes without restriction.
🧠 Processes Extensive Context: Features a 128k context window, enabling the model to understand and work within large codebases effectively.
⚙️ Utilizes Advanced Tokenizer: Employs a Tekken tokenizer with a 131k vocabulary size for efficient code processing.
Use Cases
Automated Bug Resolution in Local Projects: Integrate Devstral with a local agent scaffold like OpenHands to automatically analyze issues reported in your project's issue tracker, propose code changes across multiple files, and even test fixes, all running privately on your local machine.
Enhancing Development on Sensitive Enterprise Codebases: Deploy Devstral within your enterprise's secure environment to leverage its agentic capabilities for automating tasks like dependency updates, refactoring, or fixing identified vulnerabilities on proprietary or privacy-sensitive code that cannot be exposed to external services.
Building Advanced AI Coding Assistants: If you're developing an IDE plugin, a custom coding environment, or an internal developer tool, incorporate Devstral into your model selection to power advanced features that require understanding code context, planning multi-step solutions, and interacting with development workflows.
Conclusion
Devstral provides software engineers with a powerful, agentic AI model capable of tackling real-world coding challenges. Its strong performance on benchmarks, combined with its lightweight nature and open Apache 2.0 license, makes it a versatile tool for individual developers, enterprises, and teams building the next generation of coding tools. Whether you need local automation, secure enterprise solutions, or a robust engine for your AI assistant, Devstral offers a compelling option to enhance your software development workflow.





