Devstral Small 2

Devstral Small 2 is the small half of the coding release that Mistral AI shipped on 9 December 2025 together with the 123-billion-parameter Devstral 2 and the Mistral Vibe command-line tool. It is an agentic model rather than a chat model: it is trained to explore a repository with tools, edit several files in one task and drive a software-engineering agent through an entire fix, not to answer questions about code. The headline number is a comparison, not an absolute. On SWE-bench Verified the model scores 68.0% at 24 billion parameters — the same figure the vendor reports for GLM 4.6 at 355 billion, and within two points of Qwen 3 Coder Plus at 480 billion. The flagship 123-billion Devstral 2 reaches 72.2%, so four points of accuracy cost five times the size. On the harder Terminal Bench 2 the gap widens to ten points, which is the honest limit of the small model: it holds up on isolated repository fixes and falls behind on long terminal sessions. Technically it inherits the Mistral Small 3.1 base and the attention arrangement used in Ministral 3, ships in FP8 precision, reads images as well as text, and carries a 256,000-token context — enough to hold a mid-sized repository in view. It fits a single RTX 4090 or a 32 GB Mac. Mistral withdrew it from its own API quickly: deprecated on 27 February 2026, switched off on 31 March 2026, with Mistral Medium 3.5 named as the replacement. That has not touched its life outside the vendor. The weights are Apache 2.0 — a more permissive licence than the modified MIT covering the 123-billion flagship — and in the thirty days to early September 2026 the repository was downloaded about 230,000 times, roughly twelve times more than the flagship it was released with.

▸Videos

No videos yet.

▸Reviews