Mistral releases Devstral Small: 7B coding model for agentic tasks on consumer GPU
In one sentence Mistral releases Devstral Small, a 7-billion-parameter model fine-tuned for agentic coding that outperforms GPT-4o-mini on SWE-bench and runs on just 8GB of VRAM.
Imagine having a capable programming assistant running entirely on your own computer, no subscriptions, no internet required, and no code ever leaving your machine. Until recently, AI models that could genuinely reason about and write real-world code needed very powerful hardware, the kind found only in large data centers.
Mistral AI has changed that with Devstral Small: a 7-billion-parameter model built specifically to help developers tackle complex programming tasks. The remarkable part is that it runs on a consumer GPU with just 8GB of video memory, the kind of graphics card that many enthusiasts and professionals already own.
Devstral is not just a code autocomplete tool. It is designed for agentic tasks, meaning it can reason through multi-step problems, explore a codebase, propose solutions across multiple files, and interact with external tools. It behaves more like a junior developer colleague than a simple text suggester.
The benchmark results are impressive: on SWE-bench, one of the most respected tests for evaluating the ability to resolve real bugs in open-source projects, Devstral Small outperforms OpenAI's GPT-4o-mini, a paid cloud model that requires an internet connection.
The model is freely available on Ollama, HuggingFace, and in GGUF format, making it easy to plug into popular tools like Continue, Aider, or LM Studio. For anyone who wants to keep their code private, work fully offline, or avoid variable API costs, Devstral Small is now the benchmark for local coding models at this size class.
Companies
Mistral AI
Tools
Devstral, Ollama
Tags
Sources