What is Magistral, Mistral's first reasoning model?

Yesterday, Tuesday, June 10, 2025, the French company Mistral AI announced Magistral, its first reasoning model. Reasoning models, a category OpenAI opened with o1 last year and DeepSeek popularized with R1 in January, write out a chain of intermediate steps before giving their final answer. That extra thinking makes them better at math, logic, code and planning.

Mistral released Magistral in two versions at once, together with a research paper describing how it trained them. Its emphasis is on three things the company says early reasoning models handle poorly: depth in professional fields, transparency, and consistent reasoning in languages other than English.

What Mistral released

AIME is the American Invitational Mathematics Examination, a hard competition for high-school students often used to test reasoning models. "Majority of 64 tries" means the model answers each problem 64 times and the most common answer counts, a technique that raises scores but costs more computing. A parameter is one of the adjustable numbers inside a model; 24 billion is small enough for a single powerful machine.

In Le Chat, Mistral's chatbot, Magistral Medium powers a new Think mode. Combined with Flash Answers, a speed feature, Mistral says it produces text up to ten times faster than most competitors.

How Magistral was trained

The paper explains the method in detail. Mistral started from Mistral Medium 3, an ordinary model released in May, and taught it to reason using only reinforcement learning (RL): the model attempts problems, its answers are scored automatically, and it gradually shifts toward whatever earns higher scores. Unlike some rivals, Mistral did not copy reasoning examples from another lab's reasoning model to give it a head start. The company says RL alone improved Medium 3's single-try AIME score by nearly 50%.

The key is a verifiable reward, a score a computer can check without human judgment. For each problem the model answered, Mistral's system awarded points roughly like this:

The training algorithm, a modified version of GRPO (Group Relative Policy Optimization), has the model answer the same problem several times and compares the answers within that group. Answers that beat the group average are reinforced. Think of a teacher who grades a stack of attempts at one exercise on a curve, rather than against an outside expert's opinion.

The language reward fixes a quirk. Without it, Mistral found the model's reasoning drifted into a mix of English, Chinese and Russian. Translating a tenth of the training problems into French, Spanish, Italian, German, Chinese and Russian, and rewarding matching languages, was enough to keep the model thinking in the user's language. The authors report it then managed this even in languages it was not specifically trained on.

A worked example

Because Magistral reasons in your language and shows its steps, it suits problems where you want to check the logic. A prompt in French might look like this:

Une boulangerie vend des croissants à 1,80 € et des pains au chocolat
à 2,10 €. Samedi, elle a vendu 240 viennoiseries pour 465 €.
Combien de croissants a-t-elle vendus ? Montre ton raisonnement.

A reasoning model would set up two equations (croissants plus pains au chocolat equals 240; their prices times quantities equal 465), solve them and find 130 croissants and 110 pains au chocolat, with each step written in French that you can verify.

Why it matters

Magistral completes Mistral's lineup: the company now has a reasoning option alongside its general models, and one of the two versions is fully open. Developers can download Magistral Small, inspect how it reasons and build on it, which Mistral explicitly invites.

Mistral also aims it at regulated industries such as law, finance, healthcare and government, where a traceable chain of reasoning can help with audits. It lists uses from risk assessment and delivery planning to software design and creative writing.

TechCrunch offers a sober view: Mistral has been slower than some leading labs to ship a reasoning model, and by Mistral's own charts Magistral Medium trails Claude Opus 4 (from Anthropic) and Gemini 2.5 Pro (from Google) on several science, math and coding tests. Its distinctive selling points are speed, language coverage and openness rather than top scores. Mistral says it plans to update the models quickly.

How can you try Magistral today?

The quickest way is Le Chat, Mistral's chatbot, on the web or its mobile apps. A preview of Magistral Medium powers its Think mode; switch it on and ask a question that benefits from step-by-step reasoning. Developers can call the same preview through Mistral's La Plateforme API, or through Amazon SageMaker and IBM watsonx.

If you prefer to run a model yourself, Magistral Small is on Hugging Face under the name Magistral-Small-2506, with the permissive Apache 2.0 license that allows commercial use. You will need a computer with a capable graphics card or plenty of memory to run a 24-billion-parameter model comfortably.

Whichever version you use, give it problems with clear right answers first, compare its reasoning against your own, and remember that the benchmark figures so far come from Mistral itself.