Mistral Large 3

Mon, 01 Jan 0001 00:00:00 +0000

Mistral Large 3 is Mistral AI’s flagship, released 2 December 2025: 675 billion total parameters, 41 billion active, a 256,000-token context, and weights published under Apache 2.0.

That last fact is the reason for this page. Mistral’s documentation lists its most capable model under open weight models rather than in a premier tier — and among the eight builders in this section, no other flagship is Apache 2.0. DeepSeek publishes its current generation under MIT, which is comparably permissive; every other lab either withholds its best weights entirely or attaches conditions to them.

Mixtral 8x7B

Mon, 01 Jan 0001 00:00:00 +0000

Mixtral 8x7B, released by Mistral AI on 11 December 2023 under Apache 2.0, was the first sparse mixture of experts that large numbers of people actually ran. Sparse routing had been published years earlier; Mixtral is where it stopped being a research idea and became something a developer downloaded on a Monday.

It is on this wiki for a second reason, which is that its name is wrong in an instructive way. 8x7B reads as eight 7-billion-parameter models stapled together — 56 billion. The model has 46.7 billion parameters and uses 12.9 billion per token. Working out where those numbers come from teaches most of what a mixture of experts is.

Mistral AI on Bitsy Wiki

Mistral Large 3

Mixtral 8x7B