French AI company Mistral has released its first reasoning models and announced plans to build European cloud infrastructure, positioning itself as an alternative to American tech giants. The company unveiled two AI models called Magistral alongside a new data center project near Paris.
Magistral Small, with 24 billion parameters, is available as open-source software under the Apache 2.0 license. Users can download and modify it freely for commercial use. Magistral Medium offers more capabilities and targets enterprise customers through Mistral’s API and chat platform Le Chat.
Reasoning models work differently from standard AI systems. They think through problems step-by-step before responding, similar to OpenAI’s “o” model family and Google’s Gemini 2.5 Pro. This approach aims to improve accuracy in complex tasks like mathematics and coding. Learn more in our explainer article about Reasoning AI.
According to Mistral’s benchmarks, Magistral Medium scored 73.6% on AIME2024, a mathematics test. With majority voting, where the model generates multiple answers and selects the most common one, performance increased to 90%. However, the model underperformed compared to some competitors on certain benchmarks, including physics and programming tests.
Guillaume Lample, Mistral’s chief scientist, said the company developed its reasoning approach from scratch. Unlike competitors that often hide their thinking process, Mistral’s models show users their complete chain of thought in the user’s native language rather than defaulting to English.
Speed and multilingual capabilities
Mistral claims its reasoning models respond up to 10 times faster than competitors through new features called Think mode and Flash Answers in Le Chat. The company says responses arrive within seconds rather than minutes.
The models support reasoning in multiple languages including French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. This multilingual capability maintains high accuracy across different languages, according to the company.
During training, the models developed unexpected abilities. Magistral Medium retained image analysis capabilities even though training focused only on text-based problems. The models also learned to perform multi-step internet searches and code execution automatically.
European AI infrastructure initiative
Alongside the model release, Mistral announced Mistral Compute, a cloud infrastructure platform built with Nvidia. The project includes an initial 40-megawatt data center in Essonne, France, using 18,000 of Nvidia’s newest Grace Blackwell chips. Mistral plans to expand to 100 megawatts within 18 months.
The infrastructure aims to provide European customers with AI services that remain within EU borders and under European jurisdiction. Arthur Mensch, Mistral’s CEO, said European companies increasingly want independence from American cloud providers like Amazon Web Services, Microsoft Azure, and Google Cloud.
Mensch told the Wall Street Journal that recent political statements about US AI dominance affected European demand. He said Vice President JD Vance’s February comments about keeping America’s AI lead “tremendously affected our demand because European leaders just don’t want to be talked to that way.”
French President Emmanuel Macron supported the Nvidia partnership and called large French companies to encourage agreements with Mistral’s new data center, according to Nvidia CEO Jensen Huang. Macron described the project as “our fight for sovereignty, for strategic autonomy.”
Business performance and pricing
Mistral is on pace to earn more than $100 million in annual revenue, according to Mensch. The company raised over $600 million last year at a roughly $6 billion valuation and may seek additional funding of $1 billion or more later this year.
Magistral Medium costs $2 per million input tokens and $5 per million output tokens. This pricing is higher than Mistral’s previous models but undercuts some competitors. OpenAI’s reasoning models and Gemini 2.5 Pro charge $8 or more for output tokens.
The company targets applications requiring precision and step-by-step analysis, including legal research, financial forecasting, software development, and regulatory compliance. Mistral emphasizes that its models provide traceable reasoning processes that meet audit requirements in regulated industries.
Founded in 2023, Mistral has built its reputation on powerful open-source models while also offering proprietary enterprise versions. The company now employs about 250 people with offices in Paris, the US, and Singapore.
Sources: Mistral, TechCrunch, VentureBeat, VentureBeat, Wall Street Journal