Speed-up Kimi K3(2.8T) on a 16x GB10 Cluster — 30 t/s coding throughput, 136 t/s concurrency peak.

A user shared an update on running the Moonshot AI Kimi K3 model on a 16x GB10 cluster. The setup achieved a sustained coding generation/decode throughput of approximately 30 tok/s, with peaks around 38 tok/s during intensive code generation and agentic tasks. The cluster also demonstrated a concurrency peak of 136 t/s. Further details and resources are available on GitHub.

This report provides specific, independently verified performance metrics for the Kimi K3 model on a 16x GB10 cluster, unlike most vendor-provided benchmarks.