Mistral and All Hands drop Devstral coding models

Mistral AI just teamed up with All Hands AI to roll out the Devstral models, a lineup built specifically for agentic coding with a strong emphasis on generalizing across different prompts and scaffold frameworks.

The drop includes two distinct offerings: an open-weights upgrade named Devstral Small 1.1, and a heavier, closed-weight model called Devstral Medium.

What it does and how it works

Devstral Small 1.1 sticks to the same 24B parameter footprint as its predecessor, but it pushes coding benchmark scores significantly higher. It scores 53.6% on SWE-Bench Verified, setting a new state-of-the-art for open-source models running without test-time scaling.

Under the hood, Small 1.1 is tuned to shine alongside agent frameworks like OpenHands. It supports both native Mistral function calling and XML formatted outputs, making it easy to plug directly into different agent scaffolds and coding environments without rewriting prompts.

For harder codebases, Devstral Medium pushes the SWE-Bench Verified score up to 61.6%. Mistral claims this outperforms both Gemini 2.5 Pro and GPT 4.1, while running at roughly a quarter of their price tag.

How to get it

Devstral Small 1.1 is licensed under Apache 2.0. You can grab the model weights right off the model card to inspect, run locally, or fine-tune freely for autonomous coding setups.

Both models are also live on Mistral's API. Devstral Small 1.1 is listed as devstral-small-2507 at $0.10 per million input tokens and $0.30 per million output tokens, matching Mistral Small 3.1 pricing.

Devstral Medium is available as devstral-medium-2507 at $0.40 per million input tokens and $2.00 per million output tokens. While its weights are not public, enterprise users can deploy Medium on private infrastructure, tap into custom fine-tuning via API, or access it through Mistral Code.

Why it matters

If you are building autonomous coding agents and want full control over your stack, Devstral Small 1.1 gives you top-tier open performance in a manageable 24B package with zero commercial licensing headaches.

For teams already spending heavily on proprietary API tokens to run agentic coding loops, Devstral Medium presents a compelling trade-off: benchmark performance that holds its own against GPT 4.1 at a fraction of the operating cost.