Mistral AI, in collaboration with All Hands AI, has released Devstral, an open-source language model designed for software engineering tasks. Licensed under Apache 2.0, Devstral is tailored to tackle real-world coding challenges and is available for community use, local deployment, and enterprise applications. Here’s an overview of Devstral and its capabilities.

Addressing Real-World Coding Challenges

Unlike traditional language models that focus on isolated tasks like writing functions or code completion, Devstral is built to handle complex software development scenarios. It can process large codebases, understand relationships between components, and identify subtle bugs in intricate functions. Trained to resolve real GitHub issues, Devstral operates using agentic coding scaffolds like OpenHands and SWE-Agent, which define the interface between the model and test cases.

Devstral Performance Summary (SWE-Bench Verified)

1. Devstral: 46.8% (OpenHands scaffold)
2. Previous Open-Source SoTA: <40.8%
3. Deepseek-V3-0324 (671B): Lower than Devstral (OpenHands scaffold)
4. Qwen3 232B-A22B: Lower than Devstral (OpenHands scaffold)
5. GPT-4.1-mini: >20% lower than Devstral (custom scaffold)

Flexible Deployment for Diverse Use Cases

Devstral’s lightweight architecture enables it to run on a single RTX 4090 or a Mac with 32GB of RAM, making it suitable for local deployment on individual devices. Developers can use platforms like OpenHands to interact with local codebases and resolve issues efficiently. For enterprises with privacy-sensitive repositories, Devstral is a suitable choice for agentic coding, particularly for environments with stringent security and compliance needs.

The model can also be integrated into coding IDEs, plugins, or environments, offering developers a reliable option to enhance their toolsets. Mistral AI provides documentation and a tutorial video to guide setup and implementation. https://www.youtube.com/watch?v=oV9tAkS2Xic

Availability and Accessibility

Devstral is freely available under the Apache 2.0 license, allowing developers to build, customize, and contribute to its development. It can be downloaded from platforms like HuggingFace , Ollama , Kaggle , Unsloth, and LM Studio . For API access, Devstral is offered as “devstral-small-2505” at the same pricing as Mistral Small 3.1: $0.1 per million input tokens and $0.3 per million output tokens.

Enterprises interested in fine-tuning Devstral on private codebases or exploring advanced customization, such as continued pre-training or model distillation, can contact Mistral AI’s applied AI team for tailored solutions.

What's next

Mistral AI presents Devstral as a research preview and welcomes community feedback to improve its capabilities. The company is developing a larger agentic coding model, set to be released in the coming weeks, to further advance software engineering solutions.

This release strengthens the open-source ecosystem for coding models, and we’ll continue to monitor Devstral’s real-world performance and updates from Mistral AI.