Mistral Rejoins Open Source: Unveils High-Performance AI Model Devstral for Laptops

Mistral Rejoins the Open Source Community with the Launch of the Highly Efficient AI Model Devstral

In a significant move, French AI model manufacturer Mistral has returned to the open-source arena following criticism from parts of the community regarding its latest closed-source model, Medium3. The company has partnered with the open-source startup All Hands AI, creators of OpenDevin, to introduce a new open-source language model named Devstral. This lightweight model, featuring 24 million parameters, is specifically designed for AI software development and has demonstrated performance that surpasses many competitors with billions of parameters in certain benchmark tests.

Unlike traditional large language models (LLMs) that focus primarily on code completion or generating standalone functions, Devstral is optimized to act as a comprehensive software engineering agent. This capability allows it to understand context across files, navigate large codebases, and address real-world software development challenges. Notably, Devstral is released under the permissive Apache 2.0 license, granting developers and organizations the freedom to deploy, modify, and commercialize the model.

Mistral AI research scientist Baptiste Rozière emphasized the goal of providing the developer community with a tool that can be run locally and modified according to specific needs, with the Apache 2.0 license offering significant user flexibility.

Building on the Success of Codestral

Devstral represents the latest advancement in Mistral's code-centric model series, Codestral. Launched in May 2024, Codestral features 22 billion parameters and supports over 80 programming languages, excelling in code generation and completion tasks. The rapid iteration of Codestral has led to enhanced versions, including Codestral-Mamba and the latest Codestral 25.01, which have gained popularity among IDE plugin developers and enterprise users. The success of the Codestral series has laid a solid foundation for the development of Devstral, enabling it to expand from simple code completion to executing full proxy tasks.

Impressive Performance in SWE Benchmark Testing

In the SWE-Bench Verified benchmark tests, Devstral achieved an outstanding score of 46.8%. SWE-Bench Verified is a dataset containing 500 real GitHub issues, manually verified for accuracy. This score not only outperformed all previously released open-source models but also surpassed several closed-source models, including GPT-4.1-mini, by over 20 percentage points. Rozière proudly stated that Devstral is the best-performing open-source model to date in SWE-bench verification and code proxy tasks, and remarkably, it can run locally on a MacBook with just 24 million parameters.

Mistral AI's developer relations lead, Dr. Sophia Yang, also noted on social media that Devstral outperformed many closed-source alternatives across various frameworks. The model's exceptional performance is attributed to reinforcement learning and safety tuning techniques applied to the Mistral Small3.1 base model.

Beyond Code Generation: A Foundation for AI Software Development

Devstral's objective extends beyond mere code generation; it is designed to integrate seamlessly with proxy frameworks such as OpenHands, SWE-Agent, and OpenDevin. These frameworks enable Devstral to interact with test cases, navigate source code files, and execute multi-step tasks across projects. Rozière revealed that Devstral will be released alongside OpenDevin, which provides a scaffolding for code proxies, serving as the backend for developer models.

To ensure the model's reliability, Mistral conducted rigorous testing of Devstral across various codebases and internal workflows to prevent overfitting to the SWE-bench benchmark. They exclusively used data from non-SWE-bench datasets for training and validated the model's performance across different frameworks.

Efficient Deployment with Business-Friendly Open Source Licensing

Devstral's compact architecture of 24 million parameters allows developers to run it locally with ease, whether on machines equipped with a single RTX 4090 GPU or on Mac computers with 32GB of memory. This feature is particularly appealing for applications that prioritize privacy and require deployment on edge devices. Rozière indicated that the target users for this model include developers and enthusiasts keen on local and privatized operations, even in offline environments.

In addition to its performance and portability, Devstral's Apache 2.0 license provides significant advantages for commercial applications. This license allows unrestricted use, adaptation, and distribution, including in proprietary products, significantly lowering the barriers for enterprise adoption.

Devstral features a context window of 128,000 tokens and utilizes a tokenizer with a vocabulary of 131,000 words. It supports deployment through popular open-source platforms such as Hugging Face, Ollama, Kaggle, LM Studio, and Unsloth, and is compatible with libraries like vLLM, Transformers, and Mistral Inference.

API and Local Deployment Options

Developers can access Devstral through Mistral's Le Platforme API, with the model named devstral-small-2505, priced at $0.10 per million input tokens and $0.30 per million output tokens. For users interested in local deployment, support for frameworks like OpenHands enables immediate integration with codebases and proxy workflows. Rozière shared his own experiences using Devstral for small development tasks, such as updating package versions or modifying tokenized scripts, praising its ability to accurately locate and modify code.

Although Devstral is currently available in a research preview, Mistral and All Hands AI are already working on developing more powerful and larger follow-up models. Rozière believes that the gap between small and large models is rapidly closing, and the impressive performance of models like Devstral is now comparable to some larger competitors.

With its outstanding benchmark performance, permissive open-source licensing, and features optimized for proxy design, Devstral is not just a powerful code generation tool but is poised to become a key foundational model for building autonomous software engineering systems.

For more insights into the world of AI, stay tuned to our daily updates and explore the latest trends and innovations in AI products and applications.

Discover a wide range of innovative solutions tailored to your needs. Learn more and explore AI tools built for users on our AI Tool Directory, where you can explore features like smart search and AI assistants to find the perfect tool for you.