Mistral AI Releases New Reasoning Models

Mistral AI Introduces its Magistral Reasoning Models

The French artificial intelligence laboratory, Mistral, has entered the arena of reasoning AI models.

Announced on Tuesday, Magistral represents the company’s initial family of models specifically designed for reasoning tasks. Similar to other advanced reasoning models, such as OpenAI’s o3 and Google’s Gemini 2.5 Pro, Magistral employs a step-by-step problem-solving approach to enhance consistency and reliability in areas like mathematics and physics.

Magistral Model Variants

Magistral is offered in two distinct versions: Magistral Small and Magistral Medium.

Magistral Small, comprised of 24 billion parameters, is readily available for download via the Hugging Face AI development platform under the permissive Apache 2.0 license. (Parameters function as the internal components that govern a model’s behavior.) The more powerful Magistral Medium is currently in preview on Mistral’s Le Chat chatbot platform and through the company’s API, as well as on select partner cloud services.

According to Mistral, “[Magistral is] well-suited for a diverse range of enterprise applications, encompassing structured calculations, programmatic logic, decision trees, and rule-based systems.”

The models have been specifically fine-tuned for complex, multi-step logic, which improves interpretability and provides a clear, traceable thought process articulated in the user’s language.

Company Background and Recent Developments

Established in 2023, Mistral is a leading AI model lab focused on developing a suite of AI-powered services. These include the Le Chat platform and various mobile applications.

The company has secured substantial funding, exceeding €1.1 billion (approximately $1.24 billion), from venture capital investors such as General Catalyst.

Despite significant investment, Mistral has previously trailed some competitors in the development of reasoning models. Initial benchmarks suggest that Magistral’s release may not be highly competitive.

Performance Benchmarks

On tests evaluating physics, mathematics, and scientific reasoning – specifically GPQA Diamond and AIME – Magistral Medium demonstrates performance below that of Gemini 2.5 Pro and Anthropic’s Claude Opus 4.

Furthermore, it does not outperform Gemini 2.5 Pro on the LiveCodeBench programming benchmark.

Mistral emphasizes other strengths of Magistral in its announcement. The company claims that Magistral delivers responses at a speed ten times faster than competitors within the Le Chat environment.

It also boasts support for a broad spectrum of languages, including Italian, Arabic, Russian, and Simplified Chinese.

“Building upon our core models, Magistral is engineered for research, strategic planning, operational optimization, and data-driven decision-making,” the company states.

“This includes executing risk assessment and modeling with multiple variables, or determining optimal delivery schedules under various constraints.”

Recent Product Launches

The introduction of Magistral follows the debut of Mistral Code, a “vibe coding” client.

Prior to this, Mistral launched several models focused on coding and introduced Le Chat Enterprise, a chatbot service tailored for corporate use. Le Chat Enterprise provides tools such as an AI agent builder and integrates Mistral’s models with third-party services like Gmail and SharePoint.