Mistral Unveils Europe’s First Magistral Model: A Competitive Flop or a Game Changer?

The French AI startup Mistral has officially launched its inaugural reasoning model called Magistral. This model comes in two variants: the open-source Magistral Small, which features 24 billion parameters, and a more powerful enterprise version named Magistral Medium.

According to Mistral, the Magistral model is designed for «transparent, multilingual reasoning» aimed at professional use cases, including computation, programming logic, decision trees, and rule-based systems. In the AIME2024 test, the enterprise version achieved a score of 73.6% (where participants supported decision-making in 90% of cases across 64 questions), while Magistral Small scored 70.7% and 83.3% respectively.

This positions Magistral roughly on par with the initial Deepseek-R1 model released in January, although it still trails behind the recently updated R1-0528, as well as OpenAI’s o3 and the latest Gemini Pro 2.5 release from Google. It may seem somewhat unusual to compare Magistral with an outdated model, but Mistral likely had reasons beyond simply wanting to position Magistral favorably against its competitors.

Mistral asserts that Magistral is capable of handling «chains of reasoning» across various languages and writing systems. The company’s chat system, Le Chat, now features a «thinking mode» which, when combined with quick responses, allows for much faster answer retrieval compared to other systems.

In line with its open-source philosophy, Mistral has made Magistral Small available for use and modification by anyone.

For those interested in the technical aspects, the thesis (PDF) provides an in-depth exploration of the model’s architecture, training process, and performance metrics.

You can test the model on BotHub—a platform that doesn’t require a VPN and allows the use of Russian cards. Follow this link to claim 100,000 free tokens for initial tasks and start working!

Source