lumabri

lumabri runs huge mixture-of-experts models from a swarm of peers using the colibri engine, in pure C with no dependencies. Nothing is downloaded up front: the bytes an inference actually touches arrive from peers on first use and persist in a local mirror, with every MiB verified by sha256. Routed experts execute on the peers that hold their weights, while the local machine keeps only the dense weights, router, and KV cache. Built CPU and SSD first, it produces byte-identical output with or without a GPU, under an Apache-2.0 license.