Lumabri Lets You Run Mixture-of-Experts AI Models Across a P2P Swarm
Lumabri is a new open-source project that lets users run large Mixture-of-Experts (MoE) language models by splitting them across a peer-to-peer network of machines, using an underlying system called Colibri to handle coordination.
MoE models are built from many specialized "expert" subnetworks, but typically only a handful activate for any given query. That structure makes them a natural fit for distribution: instead of loading the entire model onto one GPU-rich server, different peers can host different experts and route requests between them.
The project appeared on Show HN with modest traction so far, but it taps into a growing interest in decentralized inference as model sizes keep outpacing what a single consumer machine can hold.
Why it matters: As frontier models grow larger, projects like this hint at a future where hobbyists and small labs pool spare hardware to run models that would otherwise require expensive cloud GPUs. It's early-stage, but P2P inference could meaningfully lower the barrier to experimenting with big open-weight models.