Magnetar

Magnetar is a large language model — that is, an AI system that understands and generates text — from the French company Mistral AI. It was released in 2025 and is designed for demanding tasks: mathematics, logic, and scientific problems. Unlike many competing products, Magnetar has been released as an open model, meaning developers can download it and run it themselves. This puts it in direct competition with models from OpenAI and Google, which can only be used via paid services.

Magnetar’s Position in the AI Competition

Most leading AI models come from the USA. Mistral AI is one of the few European companies able to compete on equal footing. Magnetar is the company’s most ambitious model to date and aims to show that Europe does not have to be technologically dependent on American or Chinese providers.

For companies and research institutions, the open format is particularly valuable. Anyone who runs a model themselves retains control over their data — it never leaves their own data center. This is especially relevant in fields such as medicine, law, or finance, where strict data protection rules apply.

How Magnetar Works Internally

Magnetar uses an architecture called Mixture of Experts. In this approach, the model consists of many specialized sub-networks, known as experts. For each query, a small control network decides which of these experts are responsible — and activates only those. The others remain idle.

This has a practical advantage: the model can be very large overall, and thus powerful, without requiring full computing power for every answer. It computes faster and more cheaply than a classical model of equal size. According to Mistral AI, Magnetar shows above-average results especially in mathematical reasoning — that is, the step-by-step solving of problems in which intermediate results are checked.

In addition, Magnetar relies on so-called reasoning, a mode in which the model engages in a kind of inner dialogue before giving its final answer. It thinks through the solution path before responding. This approach has proven significantly more reliable for difficult tasks than answering directly.

Magnetar in Practice

Technically skilled users can download Magnetar via the Hugging Face platform and run it on their own hardware. Mistral AI also offers it through its own cloud service called La Plateforme — there, users pay per request, similar to OpenAI’s API.

In practice, Magnetar mainly appears in contexts where accuracy is critical. These include automated code reviews, AI systems solving Olympiad-level math problems, or the analysis of scientific texts. Various benchmarks — standardized tests used to compare AI models — show that Magnetar matches or, in certain disciplines, even outperforms far more expensive models from OpenAI and Google on mathematical tasks.

For observers of the AI industry, Magnetar is above all a signal: a European company with a fraction of the resources of OpenAI or Google can keep up technically in certain areas. This is changing the discussion about who will control the infrastructure for AI in the future.