Pool spare GPU capacity to run LLMs at larger scale (github.com)

> MoE models via expert sharding with zero cross-node inference traffic

This makes the whole project questionable