Magistral

Magistral is Mistral AI’s first reasoning model, released on 10 June 2025. It was designed for domain-specific, transparent and multilingual reasoning — it produces an explicit, multi-step chain-of-thought that can be traced in the user’s own language.

The model comes in two variants: Magistral Small (24B parameters, open weights under the Apache 2.0 license, self-deployable) and Magistral Medium (a more capable, proprietary enterprise variant). Magistral supports reasoning in many languages, including English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese.

Magistral is available in Le Chat (a “Think” mode and “Flash Answers”), via the API on Mistral’s platform (la Plateforme, in preview), and on Amazon SageMaker (with announced support for IBM watsonx, Azure AI and Google Cloud). Magistral Small weights can be downloaded from Hugging Face. With “Flash Answers”, Magistral Medium achieves up to 10× higher token throughput in Le Chat than most competing solutions.

On the AIME 2024 math benchmark, Magistral Medium scores 73.6% (90% with majority voting maj@64), and Magistral Small 70.7% (83.3% with maj@64). The model was trained using reinforcement learning (including with verifiable rewards, RLVR), which developed its reasoning capabilities.