Exclusive: Mistral debuts its first reasoning models
Add Axios as your preferred source to
see more of our stories on Google.
/2025/06/10/1749516919386.gif?w=3840)
Illustration: Brendan Lynch/Axios
French AI startup Mistral is launching its first reasoning models, aiming to rival new offerings from OpenAI, DeepSeek and others, Axios has learned.
Why it matters: Reasoning models aim to boost accuracy by using more computing power at inference time, a trend gaining steam amid diminishing returns from building ever-larger language models.
Driving the news: Mistral is releasing two versions of its new Magistral family of reasoning models.
- The first, Magistral Small, is a 24-billion-parameter model that is being released under an open-source license.
- Magistral Medium is more powerful, but proprietary and currently available only in preview mode.
- Mistral says the new models are designed for a range of real-world tasks in areas including law, finance, health care, and engineering. The proprietary version, Mistral says, is capable of 1,000 tokens per second, outpacing the performance of rivals.
Between the lines: Unlike many models that reason primarily in English, Mistral's can "think" in the query's native language, a potentially more efficient approach.
The big picture: Based in France, Mistral has benefited from both its embrace of open source as well as the fact that it offers an alternative to the large American and Chinese tech companies.
- Just 15 months after launching its first paid products, Mistral has racked up over $100 million in contracted sales, according to a source familiar with its finances.
What they're saying: Anjney Midha from a16z, who led the company's Series A funding round and sits on its board, said that Mistral's focus on open source is its key selling point, even more than its unique geography.
- Midha noted that while big shifts in technology tend to be initially led by proprietary efforts, businesses quickly gravitate toward the advantages of open source.
- "They want cheaper, faster and more control," Midha told Axios. "It's happening now with AI. It just takes a little bit of time."
