Mistral's open-weights reasoning LLM — the European answer to o1, with a smaller footprint and a much smaller price tag.
Magistral Medium is Mistral AI's first dedicated reasoning model, trained with reinforcement learning on chain-of-thought traces to handle math, code, and multi-step logic. The Medium tier sits between Mistral's small edge models and their frontier Large series — and unlike the closed o1/o3 family, you can download the weights and self-host.
Who it's for: European enterprises with EU residency requirements, labs that need self-hostable reasoning, and developers building agent pipelines where $15/M token frontier models are too expensive at scale.
Trained with RL specifically for reasoning — breaks problems into steps, verifies intermediate results, catches its own slips. On AIME25 math benchmarks it scores ~73% vs o1's ~78%.
Weights are downloadable for commercial use. Self-host with vLLM, llama.cpp, or deploy on Azure AI Foundry. Full auditability — you can trace exactly what it's doing.
Backed by Mistral's EU infrastructure — data never leaves the continent. Critical for GDPR-sensitive enterprises and EU government contracts that can't use US-based CLOUD-only models.
$0.40/M input, $2/M output vs o1's $15/$60. For pipelines that chain 10+ reasoning calls per task, that difference decides whether the project is viable or not.
Magistral Medium is the open reasoning model to beat in 2026. If you're EU-based, regulated, or price-sensitive, it's the clear default. If you need the absolute frontier on math or agentic tool use, o1 and Claude still lead — but not by enough to ignore the 10x cost gap.
The full Mistral family — Magistral sits next to Large and Small.
Frontier Mistral — the bigger sibling when Magistral tops out.
Frontier closed model — main benchmark comparison.
Reasoning-heavy frontier for hard agentic tasks.