What's this about?
Petals — a system where many people each host a few layers of a large model and inference hops across the swarm — resurfaced on HN as open models get big enough that no single home GPU can hold them.
What exactly is "distributed inference"?
Splitting one model across many machines: like a relay race where each runner (GPU) carries the computation a few layers, then hands off. BitTorrent proved strangers will pool bandwidth; Petals bets they'll pool GPU time.
Why is it worth noting?
This week's strongest open models (Kimi K3, DeepSeek-V4 family) are trillion-parameter MoEs — interest in community-scale serving tracks directly with open-weights capability.