News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnMagistral

Magistral

Mistral AI's first reasoning model, in open and enterprise versions

Magistral is Mistral AI's first reasoning model, announced in June 2025 in two variants: an open-weight 24B Magistral Small and a stronger enterprise Magistral Medium. It is trained with reinforcement learning for step-by-step thinking, and the company says it keeps its reasoning traceable and multilingual.

Share on WhatsApp

  • Mistral AI's first reasoning model, announced in June 2025
  • Magistral Small is 24B parameters, released under Apache 2.0
  • Magistral Medium scored 73.6% on AIME2024, 90% with majority voting @64
  • Trained with a reinforcement learning pipeline built from scratch
  • Reasons in English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese

What it is

Magistral is a family of reasoning language models from Mistral AI, released in June 2025. It comes in two forms: Magistral Small, a 24-billion-parameter model whose weights are published, and Magistral Medium, a stronger version offered to enterprises. The company says it is built for problems that need several steps of deliberate thinking, such as legal research, financial forecasting, planning and coding. Its reasoning is meant to be readable in the user's own language, so a reader can follow and check each step. Mistral AI describes it as its first model of this kind and says it will keep iterating on it.

How it came about

Mistral AI announced Magistral as its first reasoning model, with a paper released on 12 June 2025 describing the training work. The team built its own reinforcement learning pipeline from scratch rather than reusing existing implementations or traces distilled from other models. The paper lists Mistral-AI as the author, with Abhinav Rastogi and 98 other contributors. The company says the release involved both experienced staff and interns, and that it wants to grow this family of models further.

Inside the model

Magistral Medium was trained for reasoning on top of Mistral Medium 3 using reinforcement learning alone. Magistral Small additionally includes cold-start data drawn from Magistral Medium. The team reports that training on text alone kept most of the starting checkpoint's abilities, and that it maintained or improved multimodal understanding, instruction following and function calling. They also present a simple way to control which language the model reasons in. The company says the pipeline let them test how far pure reinforcement learning can go with language models.

What it can do

On AIME2024, a maths competition benchmark, Magistral Medium scored 73.6%, rising to 90% with majority voting across 64 samples; Magistral Small scored 70.7% and 83.3% on the same tests. The company says the models handle structured calculation, programmatic logic, decision trees and rule-based systems, and that they work well for creative writing and storytelling. It also claims reasoning quality holds up across many languages and alphabets.

How to use it

Magistral Small is published as open weights under the Apache 2.0 licence and can be downloaded from Hugging Face for self-deployment. A preview of Magistral Medium runs in Le Chat and through the La Plateforme API. The company says Medium is also on Amazon SageMaker, with IBM WatsonX, Azure AI and Google Cloud Marketplace to follow, and that on-premises enterprise deployments are available through its sales team.

Why it matters

Mistral AI positions Magistral as a contribution to open source work on thinking models, inviting the community to inspect and build on Magistral Small. The company notes that its earlier open models have already been used for community projects such as ether0 and DeepHermes 3. It says the model is aimed at making reasoning that can be audited in high-stakes fields like law, finance, healthcare and government.

Questions people ask about Magistral

What is Magistral?

It is Mistral AI's first reasoning model, announced in June 2025. It is designed for multi-step problems where you need to see the logic behind an answer, and it comes in an open 24B version called Magistral Small and an enterprise version called Magistral Medium.

How is it different from a general-purpose model?

Mistral AI says Magistral is fine-tuned for multi-step logic, so its chain of thought is traceable and written in the user's language, rather than being a general-purpose model. The company positions it for work needing transparency and precision, such as legal research and financial forecasting.

Can I use it for free?

Magistral Small is published as open weights under the Apache 2.0 licence and can be downloaded from Hugging Face for self-deployment. A preview of Magistral Medium is available in Le Chat and through the La Plateforme API, and the company says enterprise deployments can be arranged.

What languages does it reason in?

The company says reasoning works across global languages and alphabets, and names English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese as languages it handles particularly well. The paper also describes a method for forcing the model's reasoning language.

How well does it score on benchmarks?

On AIME2024, Magistral Medium scored 73.6%, or 90% with majority voting across 64 samples. Magistral Small scored 70.7% and 83.3% on the same two tests. These are the figures Mistral AI reports for its own models.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Magistral yet, so the facts on this page come from the sources listed above: mistral.ai/news/magistral, arxiv.org/abs/2506.10910. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.