Mistral launches Europe’s first reasoning model Magistral but lags behind competitors
Update: Mistral officially announces Magistral
French AI startup Mistral has now officially introduced its first reasoning model, Magistral. The model comes in two versions: an open-source “Magistral Small” with 24 billion parameters, and a more powerful enterprise version called “Magistral Medium.”
According to Mistral, Magistral is designed for “transparent, multilingual reasoning” and is aimed at professional applications like calculations, programming logic, decision trees, and rule-based systems. On the AIME2024 benchmark, the enterprise version scored 73.6% (or 90% using a majority vote on 64 questions), while Magistral Small reached 70.7% and 83.3%, respectively.
This puts Magistral roughly on par with the first Deepseek-R1 model from January, but still well behind the recently updated R1-0528 as well as OpenAI’s o3 and Google’s latest Gemini Pro 2.5. It seems a bit odd to compare Magistral to an outdated model, but maybe Mistral had reasons for that beyond making Magistral look better in comparison to its competitors.
Ad
Mistral says Magistral can handle “chain-of-thought” reasoning across languages and writing systems. The company’s chatbot, Le Chat, now features a “Think Mode” which, in combination with Flash Answers, can reportedly produce responses much faster than other systems.
In line with their open-source approach, Mistral has made Magistral Small available for anyone to use and modify.
For those interested in the technical side, the Magistral research paper (PDF) dives into the model’s architecture, training process, and performance metrics.
Original article:
Mistral’s “Magistral”: A reasoning model that understands Europe’s languages
The French start-up Mistral is launching its first reasoning model on the market with “Magistral”. It is designed to enable logical thinking in European languages.
Recommendation
According to CEO Arthur Mensch, the model is intended to compete with established offerings such as OpenAI’s “o1” and DeepSeek’s “R1”. Magistral was presented as part of London Tech Week. The release and detailed presentation will take place later today.
Reasoning models differ from classic language models in their ability to solve complex problems logically in several steps. According to Mensch, Magistral is “great at math” and “great at programming”. Benchmark results are not yet available.
Magistral should think in European languages
A special feature of Magistral is its ability to draw logical conclusions in several European languages. “Historically, US models could think in English and Chinese models could think in Chinese,” Mensch said in an interview with CNBC. Magistral aims to close this gap and also be able to think in French, for example. The initial focus will be on European languages, but a later expansion to other languages is planned, according to Mistral. In February, the company presented Saba, a language model that was trained in Arabic and many languages of Indian origin.
According to CNBC, Mistral is also pursuing the approach of so-called open-weight models with Magistral: the training parameters of the model are to be publicly accessible and can therefore be adapted by developers, for example. This could make Magistral an interesting alternative to other open reasoning models such as Deepseek’s R1.
The start-up is supported by Microsoft. In addition to OpenAI, the technology group is also investing in Mistral and thus appears to be pursuing several AI strategies at the same time.
