CoMoE redesigns expert-parallel MoE communication for consumer GPUs, where cross-GPU traffic must pass through host memory over PCIe. Instead of using datacenter-oriented All-to-All collectives, it makes the host an active routing hub.
A host-backed multicast writes each shared token once during dispatch, while token-level host staging replaces the global synchronization barrier during output combination. On RTX 5090 GPUs, CoMoE improved throughput by up to 1.46x over SGLang, reached 85–90% of the strongest tested A800 configuration's throughput, and required at most 23.4% of its equivalent GPU investment. This makes local, privacy-preserving MoE serving more economically plausible.