Mistral, a young French startup founded in 2023, has unveiled Magistral, its first family of AI models specialized in reasoning. Similar to Gemini 2.5 Pro (Google) or Claude Opus 4 (Anthropic), these models are designed to solve complex problems by following step-by-step logic, particularly in mathematics and physics.
An AI that thinks step by step
Two variants are available: Magistral Small, with 24 billion parameters and available for free on the Hugging Face platform under the Apache 2.0 license; and Magistral Medium, a more robust version still in the testing phase, accessible via Mistral's Le Chat interface, its API, as well as via partner clouds.
In its announcement, the startup emphasizes the usefulness of its new models for concrete use cases in business: "structured calculations, programmatic logic, decision trees, and rule-based systems," with the ability to "provide an explicit trace of reasoning, in the user's language." Magistral aims to be particularly suited to multilingual environments, with claimed support for Arabic, Russian, Italian, and simplified Chinese.
Despite its ambitions, Magistral Medium struggles to compete with the best in the sector. On several benchmarks (GPQA Diamond, AIME, and LiveCodeBench), the model remains behind its American competitors. But Mistral puts forward another argument: speed. "Ten times faster than the others in Le Chat," the company assures.
The launch of Magistral is part of a broader movement to accelerate AI capabilities in Europe. During a joint event at the VivaTech trade show in Paris, Nvidia CEO Jensen Huang revealed a major project: more than 20 "AI factories" are expected to be built in Europe within two years, some even exceeding a gigawatt of power.
Among the partnerships highlighted: the one established with Mistral. The two companies will jointly create Mistral Compute, a cloud infrastructure powered by 18,000 Nvidia Grace Blackwell chips, hosted in a data center in Essonne. This sovereign cloud will then be expanded elsewhere in Europe.
This initiative has obviously received the strong support of Emmanuel Macron, who sees it as a centerpiece of France's technological autonomy strategy. "I want to encourage all large companies and startups to join this team," he declared.
Nvidia is currently pushing a global strategy: providing each country with the means to develop local AI, based on its own data. "AI starts with data. Don't export it, buy a machine instead," said Jensen Huang. And if possible, an Nvidia machine, of course! His goal is to democratize access to cutting-edge AI, including for the smallest organizations.
0 Comments