On June 10, 2025, the French AI lab Mistral announced Magistral, its first family of reasoning models, in two flavors: open-weights Magistral Small and enterprise-focused Magistral Medium. TechCrunch framed the launch plainly: Mistral was entering the same territory as OpenAI’s o3 and Google’s Gemini 2.5 Pro.
Reasoning models differ from conventional models by working through problems step by step, which buys more consistent results in domains like math and physics. The announcement opens with a claim about human thinking — “the best human thinking isn’t linear” — positioning Magistral as a model that weaves through logic, insight, and uncertainty. Mistral’s stated differentiators for Magistral are a traceable, transparent reasoning process and multilingual thinking. The company shipped a research paper alongside the models covering full evaluations, its training infrastructure, its reinforcement learning algorithm, and novel observations from training reasoning models.
Two versions: an open 24B and an enterprise model
Magistral Small is a 24-billion-parameter open-weights model, released on Hugging Face under the permissive Apache 2.0 license, so anyone can download, inspect, or build on its architecture and reasoning process. Magistral Medium is the more capable enterprise variant, available in preview on Le Chat and Mistral’s API, as well as third-party partner clouds. In the Le Chat preview, Mistral showed Magistral Medium producing a one-shot physics simulation of gravity, friction, and collisions — reasoning aimed at real-world scenarios.
The intended use cases lean toward enterprise logic: structured calculations, programmatic logic, decision trees, and rule-based systems. Mistral also framed this as the start of a fast cycle, telling users in the announcement itself to “expect the models to constantly improve.”
Benchmarks and multilingual reasoning
On benchmarks, Magistral Medium scored 73.6% on the AIME2024 math competition test, rising to 90% with majority voting at 64 samples; Magistral Small scored 70.7% and 83.3% respectively. This was Mistral’s first full benchmark disclosure for a reasoning model, with the companion paper detailing the RL training infrastructure and new observations on training reasoning models — an attempt to put the methodology in the open.
Multilingual reasoning is the other headline claim. Mistral says Magistral’s chain-of-thought keeps high fidelity across numerous languages and alphabets, naming English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. For product teams outside English-first markets, “thinks in the user’s language and leaves a checkable trail” is not marketing copy but a procurement and compliance argument: when reasoning steps appear in the business language, review gets much cheaper.
Transparent reasoning and Flash Answers
Mistral describes Magistral as purpose-built for transparent reasoning: fine-tuned for multi-step logic to improve interpretability, producing a traceable thought process in the user’s language — a property the announcement explicitly contrasts with general-purpose models. For industries that audit AI decisions, such as finance, legal, and healthcare, a verifiable reasoning chain is worth more than a few extra benchmark points.
To the standard complaint that reasoning models are slow, Mistral answered head-on: Le Chat gained a Think mode and Flash Answers, with a claimed token throughput for Magistral Medium up to 10 times that of most competitors, illustrated with a side-by-side speed comparison against ChatGPT. The pitch is that real-time reasoning and interactive feedback become feasible — speed plus a traceable process, aimed directly at the stereotype that reasoning models are unpleasant to use.
What it means for open weights and buyers
For the open-source ecosystem, Small’s permissive license means a company can deploy an auditable reasoning model on its own infrastructure and inspect how it thinks. Mistral welcomed the community to examine and improve the release, noting that its earlier open models had already been used in community projects such as ether0 and DeepHermes 3 — precedent for secondary development on open reasoning models. TechCrunch’s background note: Mistral was founded in 2023, and Magistral is its first entry in the category.
Judged from 2026, Magistral showed how a mid-size lab closes the gap to the reasoning frontier with reinforcement learning: an open model builds ecosystem and trust while the enterprise model earns revenue, with both tracks running at once. For developers at the time, the market gained one more reasoning option that was neither a Silicon Valley giant nor a black box; for the industry, the reasoning-model category formally spread beyond two or three labs to the European camp in mid-2025.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.