Mistral releases the Mistral 3 family under Apache 2.0
Originally announced by Mistral AI . We link the primary source so you can read it for yourself.
Mistral AI has released Mistral 3, a family that spans small dense models and a large mixture-of-experts, all under the Apache 2.0 licence. For a company whose earlier flagship releases carried more restrictive terms, the licensing is a significant part of the story.
What was released
According to Mistral, the family includes dense models at 3B, 8B, and 14B, alongside Mistral Large 3, a sparse mixture-of-experts the company describes as its most capable model to date. The performance claims are Mistral’s own and await independent testing, as always. What is not in doubt is the licence: Apache 2.0 across the family.
Why the licence matters
For anyone deploying models commercially, licensing is often as decisive as capability. Apache 2.0 means an organisation can download, fine-tune, and redistribute these models without seeking permission, negotiating terms, or tracking user thresholds. That removes exactly the kind of friction that some earlier open-weight licences introduced, and it puts Mistral 3 on the same permissive footing as the Qwen line.
The context
This continues a clear trend through 2026 toward permissive licensing among open-weight releases. When the strongest open models ship under Apache 2.0, the practical case for local and self-hosted deployment gets stronger, because the legal questions that once complicated commercial use largely fall away. Mistral moving its flagship family to Apache 2.0 is a notable marker of that shift.
We will add Mistral 3 entries to the model catalogue once we have verified their sizes, memory requirements, and licence terms against Mistral’s own documentation.