Lumabri is a pure‑C, zero‑dependency engine that lets any machine join a swarm of peers to run huge mixture‑of‑experts LLMs, as Napster let users share files; the design assumes CPU and SSD first, GPU only speeds up inference and never changes the output.

When a node queries the model, only the bytes actually used travel over the network, are cached locally, and subsequent queries are served from disk at full speed; the engine binary remains unchanged, and GPU support only makes it faster, never altering the output.

Quick start uses ./lumabri serve --model <path> on a node that holds the model, and ./lumabri chat --tracker <ip>:7300 --engines-dir <dir> on a client; the first answer is slower while the working set crosses the network, then the local mirror serves instantly. Missing --engines-dir can trigger a 299 GB download, and a missing key leads to an unverified model.

Commands such as /swarm and /model display the live anonymous network and available models; donors can provide disk, compute, or GB budget per model, and the tracker assigns the least‑replicated slices first. Private swarms use LUMABRI_TOKEN=S for access control. One tracker indexes many models, so a single client can switch between GLM, OLMoE, and DeepSeek without restarting the whole system.