Dynamic MoE Top-K Expert Router Gating Network
Active Experts per Token: Top-4 / 64
1M Context Window VRAM & FLOPs Summary
State: Kimi K3 Native Visual 1M Context
Total Weight VRAM
2800 GB
FP8 Precision (2.8T Total)
1M KV Cache VRAM
128 GB
1,000,000 Context Tokens
Total VRAM Allocation
2928 GB
Weights + KV Cache Combined
Active FLOPS / Token
0.35 TFLOPs
175B Active Params / Token
Weights VRAM: 2800 GB
KV Cache VRAM: 128 GB
Efficiency Gain: 2.5x compute efficiency
Architecture Export & Configuration JSON