Mistral's next sparse family exists, and nobody outside can see it
A larger "fat but sparse" MoE family entered early access with research and government partners in July — no public parameter count, no benchmarks, no ship date. For a company built on open weights, that is a notable silence.
Mistral confirmed in early July 2026 that a larger sparse mixture-of-experts family has entered early access with research and government partners. There is no public parameter count, no benchmark set, and no announced ship date.
The pattern is familiar from every other lab and new for this one. Mistral's position in the market was built on shipping real weights under permissive terms — Mistral Large 3 and Mistral Small 4 are both Apache 2.0, and the second is an unusually clever design at 119B total with roughly 6B active through 4-of-128 routing.
Early access with government partners is a different business. It is not necessarily a retreat from open weights; staged release with named partners before a public drop is a normal safety and commercial sequence. But it does mean the frontier of what this lab can do is, for now, visible only to people who signed something.
The thing worth watching is which way the eventual release goes. If the fat-sparse family ships Apache 2.0 like its predecessors, the early-access period was a runway. If it ships under a bespoke licence with usage conditions, that is the same drift the rest of the field has already made, arriving at the last major holdout.
Until then the honest summary is that a significant model exists and cannot be independently assessed by anyone.
TechTimes — Mistral AI Targets Frontier Gap With Open-Weight Model Entering July Early Access → · Sean Kim — Mistral Open-Weight Model Enters July Early Access: What Arthur Mensch Confirmed (and 4 Specs He Won't) →