Mistral releases a pair of AI reasoning models

Mistral, a pioneering French AI laboratory, has officially entered the realm of reasoning AI models with the introduction of its Magistral series. Announced on Tuesday, this new family of reasoning models is designed to tackle complex problems in a systematic manner, ensuring greater consistency and reliability in areas such as mathematics and physics. The Magistral series offers two versions: Magistral Small and Magistral Medium. The former boasts 24 billion parameters and is now available for download on the Hugging Face platform under an open Apache 2.0 license. Parameters are essential components that dictate how the model operates. Meanwhile, the more advanced Magistral Medium is currently in preview on Mistral’s Le Chat chatbot platform and through its API, as well as on selected third-party cloud services. Mistral asserts that these models are tailored for a variety of enterprise applications, ranging from structured calculations to decision-making frameworks. The blog post from the company emphasizes that the models are optimized for multi-step logic processes, which enhance interpretability and create a clear, traceable thought pattern in the user's preferred language. Founded in 2023, Mistral is focused on developing cutting-edge AI solutions, including the aforementioned Le Chat and various mobile applications. The lab has garnered significant backing from investors, raising over €1.1 billion (approximately $1.24 billion) to date. However, despite its substantial funding, Mistral has faced challenges in keeping pace with other top AI laboratories, particularly in the development of reasoning models. In terms of performance, initial benchmarks suggest that Magistral may not be as competitive as hoped. On tests such as GPQA Diamond and AIME, which assess capabilities in physics, mathematics, and science, Magistral Medium has not outperformed models like Google’s Gemini 2.5 Pro or Anthropic’s Claude Opus 4. Additionally, it fell short on LiveCodeBench, a notable programming benchmark. To highlight its advantages, Mistral claims that Magistral delivers responses at speeds that are “10 times” faster than its competitors’ offerings via Le Chat. The model also supports a diverse range of languages, including Italian, Arabic, Russian, and Simplified Chinese. The company elaborates that Magistral is engineered for tasks such as research, strategic planning, and operational efficiency, facilitating data-driven decision-making in contexts like risk assessment and modeling. The launch of Magistral follows Mistral’s recent release of Mistral Code, a coding client, and several coding-centric models, alongside the rollout of Le Chat Enterprise—a business-oriented chatbot service that integrates Mistral’s models with platforms like Gmail and SharePoint.