Tencent Paper: Pruning 25% of MoE Neurons Retains 91-98% Performance

Tencent's new paper reveals that domain capabilities in MoE models reside in cross-expert neurons. Pruning neurons by routing contribution removes 25% of parameters while retaining 91-98% of performance.

2026-11-14 ~ 2026-11-14 · 2 related posts

• Tencent paper: pruning neurons by routed contribution keeps 91-98% MoE performance — rohanpaul_ai · 2026-11-14

1 near-duplicate retellings: rohanpaul_ai