UniMoMo framework compresses MoE recommendation models via expert merging
August 8, 2026
UniMoMo uses a post-training graph coarsening approach to compress large Mixture-of-Experts (MoE) recommendation models under explicit expert budgets. It groups experts based on functional similarity using unlabeled calibration sets rather than parameter distance to minimize performance loss.
HOW THIS AFFECTS YOU
●
builderYou can deploy smaller MoE models with reduced memory footprints without retraining.
●
researcherYou can use functional similarity instead of parameter distance for expert merging.