Mistral AI Introduces Magistral: A New Multilingual Model for Advanced Reasoning

On June 10, Mistral AI unveiled Magistral, a next-generation AI system developed specifically for tasks that require structured reasoning across various professional fields and multiple languages.

Designed to meet the complex needs of industries relying on strategic decisions and operational precision, Magistral supports applications such as in-depth research, scenario modeling, and logistical planning. Whether evaluating multi-variable risks or determining delivery timelines under constraints, this model delivers accurate, explainable outputs that align with enterprise demands.

What sets Magistral apart is its tailored approach to multi-step reasoning. Mistral AI emphasizes that unlike general-purpose language models, Magistral is built to produce transparent and logically consistent responses in a wide range of languages, including English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. Its architecture is fine-tuned to ensure every step of its thought process is interpretable by humans, making it ideal for high-stakes decision-making and multilingual teams.

The model is also highly effective in software development workflows, offering notable improvements in backend logic, frontend structure, and data engineering planning—outperforming traditional models that lack specialized reasoning capabilities.

Magistral comes in two distinct editions. The open-source Magistral Small features 24 billion parameters and can be accessed via Hugging Face. For enterprise clients, Magistral Medium delivers enhanced capabilities and is currently available for testing through Le Chat, via API on La Plateforme, and on Amazon SageMaker.

To complement the launch, Mistral AI has also published a detailed technical paper outlining the methodology behind Magistral. This includes insights into the training framework, reinforcement learning strategies, infrastructure used, and the evaluation benchmarks applied to assess reasoning performance.