Mistral has unveiled its latest artificial intelligence (AI) model focused on coding, known as Devstral, in a launch on Wednesday. This open-source coding agent is designed to tackle software development tasks with the capability to address real-world engineering challenges and generate contextualized code within a codebase. Based on internal evaluations, the Paris-based company reports that this model achieved the highest score on the SWE-Verified benchmark, a notable feat for any open software engineering agent. The development of Devstral was a collaborative effort with All Hands AI.
Mistral’s Devstral Coding Agent Boasts Effective Coding Talent
The announcement, detailed in a newsroom post, outlines the functionalities of this new coding agent. In recent weeks, leading AI companies have been rolling out AI-powered coding agents. OpenAI has introduced Codex, Microsoft launched GitHub Copilot, and Google released Jules in public beta. Mistral now joins this competitive arena with its own offering, Devstral.
Mistral pointed out that while many existing open-source large language models (LLMs) can handle simple coding tasks such as writing standalone functions or completing code, they often fall short when it comes to creating contextualized code in extensive codebases. This limitation can lead to challenges in discerning relationships between various components and in detecting nuanced bugs.
According to the company, Devstral effectively addresses these issues by contextualizing coding activities within the framework of the database and existing code. Internal tests show that it achieved a score of 46.8 percent on the SWE-Verified benchmark, positioning it among the best in its category. Furthermore, it outperformed larger models like Qwen 3 and DeepSeek V3, in addition to proprietary models such as OpenAI’s GPT-4.1-mini and Anthropic’s Claude 3.5 Haiku.
On the technical side, Devstral is fine-tuned from the Mistral-Small-3.1 AI model and features an impressive context window of up to 128,000 tokens. Unlike its predecessor, it is a text-only model, lacking the vision encoder present in the Small-3.1 version. Importantly, Devstral is designed to utilize tools that allow it to navigate codebases, edit multiple files, and enhance the capabilities of other software engineering agents.
Mistral emphasized that Devstral is a lightweight model capable of operating on a single Nvidia RTX 4090 GPU or a Mac with 32GB of RAM, facilitating local deployment and on-device operation. Users interested in downloading the model can find it available on platforms such as Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio. It is distributed under a permissive Apache 2.0 license, making it suitable for both academic research and commercial applications.
Moreover, Devstral can also be accessed as an application programming interface (API). Mistral has categorized the AI agent under the name devstral-small-2505, with pricing set at $0.1 (approximately Rs. 8.6) per million input tokens and $0.3 (approximately Rs. 25) per million output tokens.
