Magistral Small 1.0

Magistral Small 1.0 is the model with which Mistral AI entered the reasoning race on 10 June 2025. It is a 23.6-billion-parameter dense model built on the Mistral Small 3.1 base, trained to write out its deliberation before the answer, and published under Apache 2.0 on the day of the announcement. Two choices made it matter beyond its benchmark line. The weights were open at a time when every strong reasoning model — from OpenAI, Google and Anthropic — was available only as a paid service. And the release came with a technical report describing the reinforcement-learning recipe used to produce the behaviour, rather than a marketing page with a table of scores. For a research group or a company wanting to build its own reasoning model, that document was worth more than the checkpoint. The capability itself was modest by later standards: 70.68% on AIME 2024 and 68.18% on GPQA Diamond, against 86.14% and 70.07% for version 1.2 fifteen months on. The sharper limitation was the context window — 40,000 tokens, which is tight for a model that spends part of that budget thinking out loud before it starts answering. Version 1.2 tripled it to 128,000. Mistral deprecated the model in its API on 31 October 2025 and switched it off on 30 November 2025, naming Mistral Small 4 as the replacement. The catalogue keeps the profile because the retirement did not end its use: in the thirty days to early September 2026 the repository was still pulled roughly 80,700 times from Hugging Face — over six times more than the technically superior final version of the same family. Founding releases accumulate tutorials, quantised forks and pipelines that keep pointing at them long after the vendor has moved on.

▸Videos

No videos yet.

▸Reviews