Devstral Small 1.1
Mistral AI · France · 2025
Seven weeks after the first Devstral, the same 24 billion parameters scored 53.6 per cent on SWE-bench Verified — and the file is still downloaded sixteen times more often than its predecessor.
Devstral Small 1.1 arrived on 10 July 2025, seven weeks after the version it replaced, on the same 23.6 billion parameters and the same Mistral Small 3.1 body. The score on SWE-bench Verified went from 46.8 to 53.6 per cent — 6.8 points on identical weights count, bought entirely in fine-tuning. The more useful improvement is one that does not fit in a benchmark table. Devstral Small 1.0 was tuned tightly around the OpenHands agent scaffold, and Mistral says plainly that the 1.1 revision generalises better to other prompts and other coding environments. A model that only works inside one harness is a component; a model that survives being moved is a tool. This release also added tool calling in the documented sense — function definitions passed with the request — which is what lets it drive an agent loop that someone else built. Everything else is inherited: a 128,000-token context, text-only input and output after the vision encoder was cut from the base model, the Tekken tokeniser with a 131,000-token vocabulary, the Apache 2.0 licence with no additional clauses, and a size that runs on one RTX 4090 or a 32 GB Mac. Mistral marked the API name devstral-small-2507 deprecated on 27 February 2026 and switched it off on 31 May 2026, naming Mistral Small 4 as the replacement. The download numbers suggest the withdrawal changed little for the people actually using it: at roughly 47,000 downloads a month it is pulled about sixteen times more often than Devstral Small 1.0, months after the hosted version stopped answering. This is the pattern that separates open weights from a hosted product — the vendor can end the service, but not the model's working life.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!