Mistral's reasoning model family — transparent chain-of-thought in 5+ languages, Apache 2.0 licensed.
Magistral is Mistral's answer to o3 and DeepSeek R1 — a reasoning model where the chain-of-thought isn't just internal machinery, it's the product. Every reasoning step is exposed in the response, in your language, inspectable and editable. Magistral Medium is the hosted workhorse; Magistral Small (24B) is the Apache 2.0 open release that put verifiable reasoning on-run wherever you can run Llama. For enterprises in regulated industries where 'the model said so' isn't an audit trail, this is the first reasoning model that shows its work by default.
Who it's for: Legal, finance, and compliance teams that need auditable reasoning; European orgs with data-residency requirements; developers who want OSS reasoning weights to fine-tune.
Unlike English-then-translate competitors, Magistral reasons natively in French, German, Spanish, Italian — critical for EU legal/compliance use.
The reasoning trace is part of the output, not hidden. Audit it, edit it, or force it down a different path mid-thought.
Magistral Small ships under Apache 2.0 — distill, fine-tune, or self-host a reasoning model with zero legal friction.
Flash Answers mode streams reasoning at ~1,000 tok/s, making chain-of-thought feel interactive instead of batch.
If your reasoning model needs to pass an audit — legal review, medical triage, financial compliance — Magistral is the only serious 2026 option. For pure benchmark-maxxing, o3 still wins. But 'shows its work, speaks your language, and you can run it yourself' is a trio no one else offers.