SlimWise boosts MoE decode throughput up to 1.81x via expert pruning

SlimWise prunes experts only during the decode phase while reusing the full prefill KV cache. On Qwen3.6-35B-A3B, 50% expert pruning raised decode throughput up to 1.81x and delivered roughly 78% end-to-end speedup with minimal accuracy loss.

• Decode-only pruning reaches up to 1.81x throughput at 50% expert pruning
• End-to-end speedup reaches about 78% on typical request patterns
• Full-model prefill KV cache is handed to the pruned decoder without conversion
• Evaluated on two MoE backbones and three pruning criteria