Mistral Large 3: a 675B sparse MoE under a true Apache 2.0 licence
Mistral Large 3 is reported as a sparse mixture-of-experts model with 41B active of 675B total parameters, a 256K context window and a true Apache 2.0 licence.
Mistral Large 3 is the French lab's current flagship, and it is reported as a sparse mixture-of-experts model with an unusually clean licence.
Reported specification
- 675 billion total parameters, of which 41 billion are active per token
- 256,000-token context window
- Apache 2.0 licence — described as a true open-source licence, not a restricted community one
The 41B-active figure is the number that matters for deployment planning. Serving cost and latency track active parameters far more closely than total parameters, so a 675B/41B sparse model behaves much more like a 41B model at inference time than its headline size suggests — while retaining the knowledge capacity of something far larger.
Why Europe's entry matters
Mistral is the only lab in the current open-weight top tier that is neither American nor Chinese. For organisations with data-residency requirements or procurement rules that favour European suppliers, that is not a marketing detail — it is sometimes the entire shortlist.
Where it fits against the field
At 256K, the context window is a fraction of Kimi K3's one million tokens, but 256K is already past the point where most real workloads stop caring — a 256K window comfortably holds a large codebase subset, a full contract set, or months of chat history.
The pairing of a genuinely permissive licence with a servable active-parameter count makes Large 3 one of the more practical choices in the tier, as opposed to one of the most impressive on paper.
