Morning Edition · Tuesday, September 29, 2026Published at 2:12 AM EDT · New York

A Training-Free Pruning Method Targets the Memory Cost of Mixture-of-Experts Models

OMP-MoE recasts expert selection as sparse signal reconstruction, choosing experts greedily in linear time and reallocating the budget across layers.

Sparse Mixture-of-Experts architectures cut the compute needed per token, but they keep every expert resident in memory, which is why a model that activates only a few billion parameters can still require several accelerators to serve it. O…