Lumabri runs huge MoE models across untrusted peers by shipping 4KB activations instead of weights.

The project (pure C, Apache-2.0, 71 stars) lazy-fetches only the bytes an inference touches and caches them locally, sending 4 KB activations to peers holding the relevant experts rather than transferring expert weights. Local and remote paths share identical code to guarantee byte-for-byte identical output. Trust is SHA256 checksums per MiB plus optional spot-check re-execution on replica peers to catch dishonest nodes. Engine binaries exist for OLMoE, GLM, Inkling, Kimi K3 and DeepSeek V4.