12.337.2025 - Mixture of Experts Powers the Most Intelligent Frontier AI Models, Runs 10x Faster on NVIDIA Blackwell NVL72

The top 10 most intelligent open-source models all use a mixture-of-experts architecture. Kimi K2 Thinking, DeepSeek-R1, Mistral Large 3 and others run 10x faster on NVIDIA GB200 NVL72. A look under the hood of virtually any frontier model today will reveal a mixture-of-experts (MoE) model architecture that mimics the efficiency of the human brain. Just Read Article - View Press Release - Visit NVIDIA Corporation