This is a live decentralized LLM — Qwen2.5-72B-Instruct running across a mesh of independent nodes, not a single server. Its weights are split across GPU nodes, so no node ever holds the complete model.

How it works: a coordinator sits at the center of the mesh and routes each request. The GPU nodes carry the model — each computes its share and hands the result to the next, so every token is produced collectively across the mesh. Around them, node clients form the wider network — smaller machines that join, route, and support the cluster. The point is the architecture: pooled, trustless compute where powerful and everyday hardware work side by side, and everyone earns for the work they do.