Mistral launches Large 4 with 1.05T parameters

Mistral unveiled its open mixture-of-experts Large 4 (ML4) model with 1.05T total and 49–52B active parameters, trained on 3,800 Nvidia Grace Blackwell GPUs. It scored 38 on the Artificial Analysis Intelligence Index — the best result outside the US and China, but behind five Chinese open models. Weights are due by the end of October.

• Artificial Analysis index: ML4 scores 38, MiMo-V2.6-Pro scores 46
• Trained on 3,800 Grace Blackwell GPUs over roughly two months
• Cost per AA task is $1.13, or $0.57 at launch pricing
• CyberGym-E2E-AA: 81.7%, the strongest result among models