Mistral AI Teases a Mystery New Flagship Model — Early Access Coming in July
On July 6, 2026, Mistral AI CEO Arthur Mensch teased a new open-weight flagship model. He described it as a new line of “fat but sparse” Mixture-of-Experts models, with early access for research institutions, government agencies, and industry partners set to begin sometime in July.
Details
- Architecture: Reported to be a MoE model expected to significantly exceed the scale of the current flagship, “Mistral Large 3” (675 billion total parameters)
- Model name: No official name has been announced. Mensch only jokingly shot down “Le Chaton Fat” as a candidate name, without revealing the actual one
- Performance and benchmarks: No specific figures have been disclosed. The model is positioned as an effort to close the “frontier gap” with leading closed API-based models
- License: Not yet confirmed, but early access is expected to be limited to research, government, and industry partners
- Timing: Early access for partners is set to begin sometime in July, with general availability planned for “summer 2026.” The exact general availability date has not been set
How to try it
- The model is currently pre-release, with early access limited to research institutions, government agencies, and industry partners
- Details such as the general availability date, official name, benchmarks, and licensing terms are still pending future announcements