Kimi K3 runs at ~4t/s on home lab with 768GB DDR5 and 2x5090

A Reddit user reports running Kimi K3 on a home setup with 768GB DDR5 and 2x5090 GPUs, achieving approximately 4 tokens per second using a custom llama.cpp fork and GGUF quantized weights.