Full Kimi K3 (2.8T params) running on 80x RTX 5090s at 20 tok/s single stream, no HBM needed

The team successfully runs the full Kimi K3 model (2.8T params) on 80 RTX 5090 gaming GPUs, a first for open weights frontier model using GDDR7 and official MXFP4 weights without HBM. Single stream 20 tok/s with room for improvement.

• RTX 5090
• open weights
• MXFP4
• Kimi K3
• 2.8T