Recently, Mistral AI, headquartered in Paris, France, in collaboration with All Hands AI, has launched a specialized open-source AI model for software development called Devstral. This model boasts an impressive parameter count of 24 billion and is currently in the "research preview" stage. It is released under the Apache2.0 open-source license, allowing developers and enterprises to use it freely for commercial purposes.
The release of Devstral marks a significant advancement in AI within the programming domain. According to Mistral AI's official blog, this model pushes the boundaries of "agentic" coding. This implies that AI not only serves as an auxiliary tool during development but can also actively engage in code generation and complex development tasks. Through this progress, developers can tackle real-world development challenges more efficiently.
Developers can access this model via the Mistral API using the specific code identifier devstral-small-2505. The cost is $0.1 per million tokens for input and $0.3 per million tokens for output. Additionally, Devstral supports downloading from platforms such as Hugging Face and Ollama, making local deployment convenient.
In performance tests, Devstral has shown strong results. Mistral AI disclosed that the model achieved a high score of 46.8% on the SWE-Bench Verified benchmark test, surpassing several leading open-source models and even some closed-source models like OpenAI’s GPT-4.1-mini. Thanks to its optimized architecture, Devstral can run smoothly on a single Nvidia RTX4090 GPU or a 32GB RAM Mac, significantly lowering the entry barrier and enabling more developers to get started easily.
Mistral AI stated that Devstral excels particularly in codebase exploration and multi-file editing and is compatible with code agent frameworks such as OpenHands and SWE-Agent, effectively addressing real GitHub issues. This means that both novice developers and seasoned engineers can benefit from it and apply it to various development projects.