EchoMoE wants GPUs to stop repeating themselves. In a November 18, 2026 announcement, developer @fyliufengyuan reported that the technique cut inter-node communication by 58% and raised inference throughput by 22.6% on Qwen3-30B-A3B. The developer also reported a 3.8-percentage-point average decline across 18 benchmarks. Source Source
EchoMoE targets the time spent shipping activationsβthe modelβs internal representationsβbetween GPUs during mixture-of-experts (MoE) inference. Source
According to the developerβs explanation, EchoMoE caches an earlier hidden state as a reference. Subsequent layers transmit a compressed difference for expert dispatch, the step that sends representations to experts for processing. EchoMoE also compresses expert outputs during the combine step and refreshes the reference periodically. Source
@fyliufengyuan also says EchoMoE can adapt existing pretrained MoEs without pretraining from scratch. For Qwen3-30B-A3B, the developer reports that adaptation requires approximately 8,500 times fewer training tokens than pretraining. That still involves adaptation training; this isnβt a training-free switch. Source
The text of @fyliufengyuanβs results post doesnβt identify the hardware, baseline configuration, absolute throughput, benchmark names or per-benchmark scores. Without those details, the reported gains donβt establish performance in another setup, and the average benchmark decline doesnβt reveal which capabilities took the hit. Source
Comments