Devstral Small 1.0

Devstral Small 1.0 was published on 21 May 2025 under the Apache 2.0 licence, built with the agent company All Hands AI. It was not a general chatbot: the model was tuned to work inside a coding agent — to explore a repository with tools, edit several files at once and carry a task through to a patch. The headline number was 46.8 per cent on SWE-bench Verified, measured in the OpenHands scaffold, six points above the previous best open-source result. The comparison table the vendor published is the interesting part: the same benchmark gave Claude 3.5 Haiku 40.6 per cent and GPT-4.1-mini 23.6 per cent, and Mistral claims the model also beats far larger open competitors such as DeepSeek-V3-0324 and Qwen3 235B-A22B when all of them run in the same scaffold. Agentic coding, on this evidence, rewards fine-tuning for the harness more than raw parameter count. The body is Mistral Small 3.1 with the vision encoder cut out — Devstral is text-only by design — leaving 23.6 billion parameters, a 128,000-token context and the Tekken tokeniser with a 131,000-token vocabulary. At that size it runs on a single RTX 4090 or a Mac with 32 GB of memory, which is the whole point of the release: a coding agent that does not have to send the repository to anyone. Mistral withdrew the hosted version quickly. The API name devstral-small-2505 was marked deprecated on 31 October 2025 and switched off on 30 November 2025, with Mistral Medium 3.5 named as the replacement — a general model, not another coding one. The weights remain downloadable, but at roughly 2,900 downloads a month they have been overtaken by the 1.1 revision that followed seven weeks later.

▸Videos

No videos yet.

▸Reviews