Mistral drops Devstral: A local-friendly coding model that's actually open
Mistral just unveiled Devstral, a 24B coding model built with All Hands AI under a fully commercial Apache 2.0 license. It fits on a single RTX 4090 or a 32GB Mac, beating Google's Gemma 3 27B and DeepSeek's V3 on SWE-Bench Verified. Why it matters: Developers finally get a capable, agent-ready code model they can run locally without restrictive licenses.
Devstral is a new 24-billion-parameter open-source model built specifically for software engineering, developed jointly by Mistral and All Hands AI. It is designed to act as the reasoning engine behind coding agents right on your local machine.
Unlike traditional autocomplete tools, Devstral is tuned to explore entire codebases using external tools, edit across multiple files, and plug straight into agent scaffolding like OpenHands or SWE-Agent. Mistral claims it outperforms prominent open weights like Google’s Gemma 3 27B and DeepSeek’s V3 on the SWE-Bench Verified benchmark.
Local hardware specs and licensing
Despite its strong benchmark numbers, Devstral is intentionally compact. It is light enough to run on a single consumer-grade Nvidia RTX 4090 GPU or an Apple Mac with 32GB of unified memory, making friction-free local inference genuinely practical.
Crucially, Mistral licensed Devstral under Apache 2.0. This fixes the biggest pain point of its predecessor, Codestral, whose restrictive license expressly banned internal commercial deployment. With Devstral, teams can integrate, fine-tune, and ship products without commercial limitations.
Availability
Mistral is releasing Devstral as a research preview. You can pull the weights directly from platforms like Hugging Face for local hosting, or access it via Mistral's API priced at $0.10 per million input tokens and $0.30 per million output tokens. Mistral noted that a larger agentic coding model is also slated to drop in the coming weeks.
Why it matters
If you care about keeping proprietary code on-prem, Devstral hits a sweet spot that few open models manage: it fits on a single workstation GPU while offering genuine multi-file agentic capabilities.
Freeing the model under Apache 2.0 removes the legal landmines that held Codestral back from serious commercial adoption. Anyone running a 32GB Mac or a single modern GPU should give this a spin paired with OpenHands before paying for another proprietary subscription.
Sources (1)
- techcrunch.com(opens in a new tab)Originaltechcrunch.com