Mistral AI releases Magistral, its first reasoning models

Background

A reasoning model is a language model trained to write out a chain of intermediate steps before it gives a final answer. Rather than replying in a single pass, it spends extra tokens working through a problem, which helps on tasks where the answer depends on several connected steps, such as competition mathematics. The cost is that the model produces far more text per question, so the speed at which tokens are generated becomes part of how usable the model is.

What was released

On 10 June 2025 Mistral AI released Magistral Small and Magistral Medium, its first reasoning models ^[1]. The release came in two versions. Magistral Small has 24B parameters and was published under the Apache 2.0 licence ^[1], a permissive open-source licence that allows the software to be reused and modified by others. Magistral Medium was offered as a preview: it could be tried in Le Chat, Mistral's chat interface, or called through the API on La Plateforme ^[1].

Benchmark results

Mistral reported that Magistral Medium scored 73.6% on AIME 2024 ^[1]. AIME is an invitational mathematics examination for high-school students in the United States, and its problems are commonly used as a benchmark for testing how well a model handles multi-step mathematical reasoning. With majority voting, the score rose to 90% ^[1]. Majority voting means the model is asked the same question several times and the most frequent answer among the samples is taken as the final one, a technique that trades more computation for a higher chance of landing on the right result.

The models were also built to reason in eight languages, including Arabic, Russian and Simplified Chinese ^[1]. Multilingual reasoning is harder than multilingual conversation, because the chain of thought itself has to stay coherent in the language being used.

Mistral said that Magistral Medium achieves up to 10x faster token throughput than most competitors ^[1]. Throughput is the number of tokens a model generates per second. It matters more for reasoning models than for ordinary chat models, since a long chain of thought multiplies the amount of output produced for every question.

The accompanying paper

The release was accompanied by a research paper on the reinforcement learning training behind the models ^[2]. Reinforcement learning is a training method in which a model is rewarded for outputs that meet some criterion and adjusted towards producing more of them, rather than only imitating examples. The arXiv paper titled "Magistral" (2506.10910) was submitted on 12 June 2025 and describes a scalable reinforcement learning pipeline ^[2].

Why it matters

Magistral marked Mistral's entry into reasoning models, a category that had become distinct from ordinary chat models. The company shipped two versions at once: a smaller model released under a permissive licence that others can build on, and a larger model that Mistral reported reaching a high score on a widely used mathematics benchmark ^[1]. The stated throughput advantage addresses one of the practical drawbacks of reasoning models, namely that the extra thinking steps make responses slower and more expensive to serve ^[1]. The multilingual support and the published description of the training pipeline together made the release as much about how such models are built as about the models themselves ^[2].