What we know

A homelabber used a second-generation Intel Optane Persistent Memory (PMem) 200 server combined with two Nvidia RTX 3090 GPUs to run Moonshot AI’s Kimi K3 model, which has 2.8 trillion parameters. This setup achieved a processing speed of 3.11 tokens per second. The system also incorporated a RAM drive to support the workload.

Why it matters

Kimi K3 is a very large language model, and running it typically demands extensive hardware resources usually found in enterprise environments. This demonstration shows that with innovative configurations—such as combining Optane PMem with high-end consumer GPUs—homelab setups can handle models of this scale. Intel’s Optane PMem offers high-capacity, low-latency memory that can help bridge the gap between traditional RAM and storage, while Nvidia’s RTX 3090 GPUs provide significant AI processing power for such tasks.

What is still unknown

No independent verification of the performance or accuracy of this setup has been reported. Details about supported languages, broader availability of the Kimi K3 model, and any planned software updates or patches remain unclear.