Mistral Releases Devstral, an Open-Source Agentic Coding AI Model That Outperforms GPT-4.1 Mini

Mistral released a new coding-focused artificial intelligence (AI) model on Wednesday. Dubbed Devstral, it is an open-source coding agent that can perform software development tasks. The Paris-based AI firm says that, unlike other open software engineering (SWE) agents, this model can solve real-world software engineering problems and write contextualised code in a codebase. The company claimed that the AI model ranked at the top of the SWE-Verified benchmark based on internal tests. Notably, Mistral built the AI model in collaboration with All Hands AI.

Mistral's Devstral Coding Agent Is Claimed to Offer Practical Coding Capabilities

In a newsroom post, the French AI firm announced the coding and detailed its capabilities. In the last month, major AI players have been focusing on bringing AI-powered coding agents to the market. OpenAI introduced Codex, Microsoft unveiled the GitHub Copilot coding agent, and Google made Jules available as a public beta. Now, Mistral has also joined these companies with Devstral.

Mistral said that while existing open-source large language models (LLMs) are capable of performing isolated coding tasks such as writing standalone functions or code completion, they struggle when it comes to writing contextual code in a large codebase. The company added that this leads to issues when AI agents have to identify relationships between disparate components and also identify subtle bugs.

The company believes Devstral solves this problem with its ability to contextualise coding tasks with the database and existing frameworks. Based on internal testing, it scored 46.8 percent on the SWE-Verified benchmark, ranking in the top position. It also outperformed larger open-source models such as Qwen 3 and DeepSeek V3, as well as proprietary models such as OpenAI's GPT-4.1-mini and Anthropic's Claude 3.5 Haiku.

Coming to the architecture, Devstral is fine-tuned from the Mistral-Small-3.1 AI model and has a context window of up to 1,28,000 tokens. It is a text-only model as it comes without the vision encoder that was equipped in the Small-3.1. Notably, Devstral can use tools to explore codebases, edit multiple files, and power other SWE agents.

Mistral says that Devstral is a lightweight model that can run on a single Nvidia RTX 4090 GPU or a Mac with 32GB RAM. Due to this, it can be deployed locally and run entirely on-device. Those who wish to download the model can access it from Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio. It is available with a permissive Apache 2.0 licence that allows for both academic and commercial usages.

Additionally, it can also be used as an application programming interface (API). Mistral has listed the AI agent under the name devstral-small-2505, and it costs $0.1 (roughly Rs. 8.6) per million of input tokens and $0.3 (roughly Rs. 25) per million of output tokens.