Mistral AI today announced Magistral, its first reasoning model, releasing it in two variants: an open-source 24B parameter Small version under Apache 2.0, and a more powerful Medium enterprise edition. The model is designed for transparent, multi-step reasoning across domains including legal, finance, healthcare, and software engineering, with native chain-of-thought support in languages such as English, French, Spanish, Arabic, and Simplified Chinese.

On the AIME 2024 benchmark, Magistral Medium scores 73.6% (90% with majority voting @64), while Magistral Small scores 70.7% (83.3% with voting). Mistral claims its Flash Answers feature in Le Chat delivers up to 10x faster token throughput than competitors. The release is accompanied by a paper detailing training infrastructure, reinforcement learning algorithm, and observations on training reasoning models.

On Hacker News, the announcement drew 941 points and 424 comments. Users noted that the benchmarks compare against older DeepSeek models rather than the more recent R1-0528, while others questioned the real-world behavior of the model, with one commenter observing that the small model seemed ‘overcooked’ by RL training, producing boxed formatting even in non-math contexts. The debate over whether LLMs can be said to ‘think’ or ‘reason’ also flared up again in the thread.

Community member danielhanchen quickly posted GGUF quantizations, and users began sharing early impressions. The release marks Mistral’s entry into the reasoning model race, in a field that already includes reasoning models from DeepSeek.