Magistral Medium 1.2

Magistral Medium 1.2, released on 18 September 2025, was the strongest reasoning model Mistral AI ever sold. The vendor's own table puts it at 91.82% on AIME 2024 and 83.48% on AIME 2025, 76.26% on GPQA Diamond and 75.00% on LiveCodeBench v5 — a jump of roughly twenty points over version 1.1 on the mathematics sets, achieved in three months. It shared the September upgrade with its small sibling: a vision encoder, so reasoning could run over images, and a 128,000-token context. What it did not share was the licence. Magistral Small went out under Apache 2.0 as a downloadable file; Medium was a Premier model, available only through the vendor's API and its cloud partners, with the weights never published. That difference decided what each model is today. Mistral deprecated Medium 1.2 on 22 May 2026 and switched it off on 31 July 2026, pointing users to Mistral Medium 3.5 — a general model, not a reasoning one. For the open siblings retirement meant the end of a service: the files are still on Hugging Face and still being downloaded in five- and six-figure volumes. For Medium 1.2 it meant the end of the model. No copy exists outside the vendor's infrastructure, no one can reproduce a result obtained with it, and the profile below is, together with the benchmark tables the lab published, most of what remains. The catalogue keeps the entry for that reason. A model that scored above 90% on a hard mathematics benchmark, and lasted ten months, is part of the record of how quickly this class of system was built and discarded.

▸Videos

No videos yet.

▸Reviews