Physix FrontierPhysix Frontier · News Briefing Card (TechCrunch · Oct 6, 2026)
Mistral launches 1T-parameter multimodal model Mistral Large 4
KEY FACTS
- French AI company Mistral AI released its multimodal large model Mistral Large 4 on Tuesday, with a parameter count of 1 trillion.
- The model's weights are not yet open; it is accessible only through a public secure endpoint, with weights planned to open in three weeks.
- Mistral says ML4 was trained entirely on its own compute, using only 4,000 Nvidia GPUs.
- The company claims its GPU usage is two to three times lower than that of Chinese competitors, and also notably lower than that of closed-source rivals.
- ML4 focuses on optimizing scenarios such as cybersecurity, finance, and chip design; benchmark results have not yet been published.
KEY DATA
1 trillionParameter count
4,000 Nvidia GPUsTraining GPU count
€21 billion (about $24.39 billion)Series D valuation
3 weeksOpen-weights wait time
PHYSIX OBSERVATION
Mistral trained a trillion-parameter model with 4,000 GPUs. If true, it shows training efficiency is becoming a more critical competitive dimension than piling on compute. But with weights delayed by three weeks and benchmarks not yet out, the so-called "third path" currently looks more like a posture. Whether Europe can stand firm between open source and closed source will depend on the community's actual scrutiny once the model is opened.
Source: TechCrunch report