Magistral

What it is

Magistral is a family of reasoning models from Mistral AI. A reasoning model works through intermediate steps before producing a final answer, rather than only returning a direct response. Magistral was Mistral AI's first family of this kind, released on 10 June 2025. It launched with two members: the open-weight Magistral Small and the enterprise Magistral Medium ^[1]. Small has 24B parameters and was released under Apache 2.0, while Medium was the enterprise model in the launch pair ^[1]. The family therefore covered an openly licensed model and an enterprise model at the same release ^[1].

Launch and training

Magistral Medium was available in preview in Le Chat and through the API ^[1]. In that offering, Flash Answers gave up to 10x faster token throughput ^[1]. A paper describing the reinforcement learning training was posted to arXiv as 2506.10910 ^[2]. The models reason in English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese ^[1]. The documented set includes several major languages and makes multilingual reasoning part of the model behavior described at launch ^[1].

Benchmark results

On AIME 2024, Magistral Medium scored 73.6%, and 90% with majority voting at 64 samples ^[1]. Magistral Small scored 70.7%, and 83.3% with majority voting ^[1]. AIME 2024 is a mathematics benchmark, so these numbers describe performance on mathematical problems. Majority voting means generating multiple candidate answers and selecting the most common one; the reported scores show that this procedure raised the results for both models ^[1]. The benchmark numbers are the concrete performance figures attached to the launch models ^[1].

Updates and retirement

Magistral 1.1 was released on 24 July 2025, and Magistral 1.2 on 17 September 2025 ^[3]. Mistral's changelog lists Magistral Medium 1.1 and Magistral Small 1.1, with identifiers magistral-medium-2507 and magistral-small-2507, as released on 24 July ^[3]. The 1.2 models were retired on 31 July 2026 and folded into Mistral Small 4 and Mistral Medium 3.5 ^[4]. Retirement details were listed per model. Magistral Medium 1.2, identifier magistral-medium-2509, had a deprecation date of 22 May 2026 and a retirement date of 31 July 2026, with Mistral Medium 3.5 as its alternative ^[4]. Magistral Small 1.2, identifier magistral-small-2509, had a deprecation date of 30 April 2026 and a retirement date of 31 July 2026, with Mistral Small 4 as its alternative ^[4].

Where it sits

Magistral marks Mistral AI's first reasoning-model family, with an open-weight model and an enterprise model launched together ^[1]. Its documentation covers multilingual reasoning and a paper on reinforcement learning training ^[1][2]. The version history runs from the initial release to 1.1 and 1.2, then to retirement and replacement by Mistral Small 4 and Mistral Medium 3.5 ^[3][4]. The name therefore refers to a changing family: initial models, updated versions, and eventually retired versions ^[3][4].