Moonshot has released the weights for its Kimi K3 model on Hugging Face, featuring 2.8 trillion parameters and a mixture-of-experts architecture with 16 active experts per token. The team plans to deploy the model across A100, H200, and B300 GPUs, with benchmark results expected this week.

  • Kimi K3 utilizes quantization-aware training in MXFP4, resulting in approximately 1.4 TB of weights.
  • Deployment on 8x A100s requires three nodes due to memory constraints and lacks native FP4 support.
  • An 8x H200 setup also requires at least two nodes, incurring interconnect costs.
  • Only the 8x B300 configuration fits the entire model in a single node with room for KV cache.

The authors will provide tokens per second, time to first token, and cost metrics for all three GPU configurations by the end of the week.