French artificial intelligence laboratory Mistral has unveiled its new Magistral series of inference models, designed to boost performance in logical reasoning tasks across disciplines like mathematics and physics. The release marks a significant step in Mistral's push to compete with larger AI developers.

The Magistral lineup includes two variants: Magistral Small (24 billion parameters) and Magistral Medium. The Small version is currently available for free download on Hugging Face under the Apache2.0 license, while the more powerful Medium variant remains in preview mode, accessible only through Mistral's Le Chat platform and partner APIs.

According to Mistral's technical blog, these models employ a step-by-step problem-solving approach to improve consistency and reliability in outputs. The company highlights their suitability for enterprise applications involving structured calculations, programmatic logic, and rule-based systems. Through specialized fine-tuning, the models provide enhanced explainability by displaying traceable thought processes in the user's preferred language.

Founded just last year, Mistral has quickly attracted substantial backing—including over €1.1 billion from investors like General Catalyst—but still trails industry leaders in model development. Benchmark tests reveal that Magistral Medium underperforms against competitors like Gemini 2.5 Pro and Claude Opus 4 in physics and mathematics evaluations (GPQA Diamond and AIME tests), as well as programming assessments (LiveCodeBench).

However, Mistral emphasizes distinct advantages in processing speed and multilingual support. The company claims Magistral responds ten times faster than rival models on its Le Chat platform while handling languages including Italian, Arabic, Russian, and Simplified Chinese.

The launch follows closely on Mistral's recent introduction of coding-focused tools like the "atmosphere programming" client Mistral Code and Le Chat Enterprise—a business-oriented service integrating with platforms like Gmail and SharePoint. These strategic moves position Magistral as a solution for research, operational optimization, and data-driven decision making involving complex calculations or multi-factor risk assessments.

Key Points

1. Mistral's new Magistral models specialize in step-by-step logical reasoning for technical disciplines
2. The Small version (24B parameters) is freely available on Hugging Face; Medium remains in limited preview
3. Benchmarks show performance gaps versus leading competitors but advantages in speed (10x faster responses) and multilingual support
4. Models target enterprise applications including structured calculations and rule-based systems
5. Release continues Mistral's rapid product rollout following recent coding tools and business services