Mistral AI Unveils Magistral: A Groundbreaking Model Designed to Think Like a Human!

This spring of 2025, Mistral AI is set to make a significant impact. Over the past few weeks, the French company has unveiled several new initiatives including the Small 3.1 language model, a business-oriented plan, an API for crafting personalized agents, and a coding-specific agent. On Tuesday, June 10, 2025, the startup took another major step by launching its first reasoning model, aptly named Magistral. Here’s a detailed look at its features.

Mistral AI Releases its “First Reasoning Model”

While not their first launch, Magistral marks a departure from Mistral AI’s generalist models such as Mistral 7B, Mixtral, and Small 3.1. The company notes that the field is still in its infancy: “The lack of specialized depth needed for domain-specific problems, limited transparency, and inconsistent reasoning in the desired language are just a few of the shortcomings of earlier thought models.”

Magistral distinguishes itself by specializing in complex and structured reasoning. It is designed to simulate logical, deliberate, and traceable thought processes akin to human reasoning, including the progression of thought steps, especially in demanding contexts. It is based on the ‘chain-of-thought’ principle, which has become prevalent in AI models, notably favored by OpenAI.

Optimized for multi-step logic, Magistral enhances interpretability and provides a traceable thought process in the user’s language, unlike general-purpose models, Mistral AI specifies.

Magistral is released in two editions: Magistral Small, which is open source, and Magistral Medium, a more powerful version aimed at businesses.

Multilingual, Fast, and Versatile

Mistral AI emphasizes the strengths of its reasoning model. Magistral showcases the following features:

• Mathematical Logic Performance: Magistral Medium scored 73.6% on AIME2024 and up to 90% using a majority vote system, while Magistral Small reached 70.7% and 83.3% under the same conditions.
• Advanced Multilingual Reasoning: Magistral employs clear and structured logic across a variety of languages and alphabets, showing high performance in English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese.
• Professional Versatility: Magistral adapts to a broad range of contexts from structured logical computations and programmatic tasks to legal research, financial forecasting, or software development. It is equally suitable for corporate strategies and regulated sectors.
• Accelerated Responses with Think Mode and Flash Answers: Within the Le Chat interface, these features enable responses up to 10 times faster than other models while maintaining reasoning quality. However, this feature is available only in paid plans.

Mistral AI: How to Test Magistral Models

The Magistral Small model is available for self-deployment under the Apache 2.0 license. You can download it from the link provided below.

Download Magistral Small

On the other hand, Magistral Medium can be tested using Le Chat, Mistral AI’s conversational agent, as well as through its La Plateforme interface.