French artificial intelligence startup Mistral AI has unveiled its first large-scale thinking language model, Magistral, which is also the first of its kind developed in Europe, as the startup tries to keep up in a market dominated by Chinese and American companies.
Magistral, as Mistral AI says in the announcement, “has a perfect command of transparent and multilingual reasoning focused on a specific field of activity.” The new model will be available in open and enterprise versions, among other things.
Users will be able to choose between the Magistral Small, an open-source version with 24 billion parameters, and the Magistral Medium, a more powerful enterprise version. Test results show that the Mistral models show results similar to the Chinese DeepSeek R1 and V3 models.
The startup notes that the new models work in different languages and alphabets of the world and are suitable for a wide range of corporate tasks - from structured calculations and program logic to decision trees and rule-based systems.
In addition, the capabilities of the models will be available in the Mistral AI chatbot – Le Chat. Thanks to the new Think mode, users will be able to receive answers ten times faster than from many competing chatbots.
Mistral AI also invites the community to join in modifying and developing the open Magistral Small model.
"Because Magistral Small is open source, we invite the community to explore, modify, and evolve its architecture and reasoning processes to further accelerate the emergence of language models of thinking. Our previous open models have already been used by the community for such exciting projects as ether0 and DeepHermes 3," the company said in a statement.
We remind you that the French Mistral AI is perhaps the brightest representative of Europe in the field of artificial intelligence. According to a recent assessment by venture capitalists, the startup is valued at $6.2 billion. Currently, the French are taking active steps in the development of AI infrastructure in Europe, in particular, thanks to an agreement with NVIDIA to build data centers necessary for training artificial intelligence.