MoE expert prediction achieves 150-200 tokens/sec offload

A user experimented with predicting which MoE experts are used next token to reduce idle GPU time during CPU/GPU offload on a 3060 12GB, boosting throughput from 30 tg/s to 150-200 tg/s. The approach uses Qwen3.6 35B A3B GGUF and shows potential for significant speedups in local inference, though practical implementation remains unconfirmed.