Mistral, a Paris‑based AI startup founded in 2023, has entered the reasoning AI space this week with the launch of its first model family called Magistral. Like OpenAI’s o3 and Google’s Gemini 2.5 Pro, the new models are designed to work through problems step by step, which is ideal for tasks in maths, physics, or logic. Magistral comes in two versions. Magistral Small offers 24 billion parameters and is available to download openly under the Apache 2.0 license via Hugging Face. Magistral Medium is more powerful but available only in preview through Mistral’s Le Chat chatbot or a select API and cloud partners.
This launch reflects Mistral’s ambition to reach enterprise users who need reliable reasoning. Though founder funding exceeds €1 billion, Mistral has lagged behind rivals in this niche. The company highlights that Magistral Medium is ten times faster in its Le Chat environment and supports many languages, including Italian, Arabic, Russian, and Chinese. Such speed and multilingual support could play well with global developers and businesses facing time-sensitive challenges.
Performance remains mixed. In tests such as GPQA Diamond, AIME, and LiveCodeBench, Magistral Medium did not match models like Gemini 2.5 Pro or Anthropic’s Claude Opus 4. However, early results from VentureBeat show better scores when majority voting is used, yielding around 90 percent on AIME via the voting method. Mistral also experimented with reinforcement learning from open datasets, which further improved performance by adding 5 to 12 percentage points on key benchmarks.
Mistral will likely market Magistral for use cases requiring clear reasoning trails, such as legal research, risk modelling, or optimisation problems. Its traceable chain of thought makes the model useful for compliance tasks where answers must be auditable. At the same time, offering the Small version openly encourages experimentation in research and developer communities, helping to build visibility while it refines its tech.
By joining the reasoning model race, Mistral shifts from a general LLM focus to a specialised capability space. While Magistral Medium may not yet lead benchmarks, its speed and language breadth, combined with enterprise reach, give it a niche. For Mistral, this feels like a coming of age, a step that brings it closer to real-world reasoning applications and shows promise in a crowded field.
