[Breaking] Kimi K3 deployed on 8x NVIDIA B300 on Day 0: 2.8T MoE runs on single node without quantization

On the same day weights were released, Kimi K3 (2.8T MoE) was deployed on a single node with 8x NVIDIA B300, running without extra quantization. Deployment method, GPU memory usage, inference and coding performance are reported.

• benchmark
• MoE
• deployment
• NVIDIA B300
• Kimi K3