AnalysisAI ModelsSeptember 20, 2026

Kimi K3 (2.8T) runs on 16x GB10 cluster at 30 t/s coding

A user reports running the full Moonshot AI Kimi K3 (moonshotai/Kimi-K3) 2.8T-parameter model across a 16x GB10 cluster, hitting 30 t/s coding throughput and a 136 t/s concurrency peak. Custom runtime patches and a tuned network layout were needed to get the model running smoothly.

1 source

More stories today

Open the live feed