Mistral has released Devstral, an open-source AI agent designed specifically for software engineering tasks across complex, real-world codebases.
Highlights
Developed in collaboration with All Hands AI, Devstral aims to bridge the gap left by many large language models that often underperform when required to work contextually across multiple files in large software projects.
Unlike traditional code models focused on isolated tasks such as function completion or code snippet generation, this one is designed to act as an agentic model—capable of understanding and reasoning over entire codebases.
It can edit multiple files, navigate interdependent components, and perform contextual code reasoning in a way that reflects real-world developer workflows.
Performance on SWE-Verified Benchmark
Devstral’s effectiveness is demonstrated through its score of 46.8% on the SWE-Verified benchmark, a test designed to assess performance on complex software engineering tasks.
This result places it at the top of the leaderboard, outperforming both proprietary and open-source models, including OpenAI’s GPT-4.1 Mini, Qwen 3, Anthropic’s Claude 3.5 Haiku, and DeepSeek V3.
Efficient Architecture for Local Deployment
The model is built on Mistral-Small-3.1, a compact yet high-performing architecture with 24 billion parameters and a 128,000-token context window.
It is a text-only model, omitting the vision encoder included in the original Small-3.1, to focus on optimizing token efficiency for code-related tasks. This design allows Devstral to run efficiently on local hardware such as a single NVIDIA RTX 4090 GPU or a Mac with 32GB RAM.
This lightweight architecture enables individual developers and small teams to deploy Devstral without relying on cloud infrastructure, supporting workflows that prioritize speed, control, and privacy.
Integration and Use Cases
In addition to being a standalone agent, Devstral is built for integration with other tools in a software development pipeline. It supports tool usage, code exploration, and file editing—making it more versatile than traditional completion-based models.
It can also power larger systems that require agentic behavior across code environments.
Devstral is part of Mistral’s broader ecosystem, complementing other task-specific models such as Codestral for code generation and Mathstral for mathematical problem-solving. This modular strategy allows developers to choose specialized tools optimized for different domains.
Licensing and Accessibility
Released under the Apache 2.0 license, Devstral is available for both academic research and commercial applications. It can be accessed through popular platforms including Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio.
Developers looking to use the model via API can do so under the identifier devstral-small-2505.
Pricing for API usage is set at $0.10 (approx. Rs 8.6) per million input tokens and $0.30 (Rs 25) per million output tokens, positioning Devstral as a cost-effective solution for high-performance, local AI-assisted coding.
While proprietary models like GPT-4.1 Mini remain widely used, Devstral introduces an open alternative that offers competitive performance and lower operational overhead.
Its local deployment capability adds flexibility and security for teams seeking alternatives to cloud-based LLM platforms.
