Tag
PuzzleMoE introduces a pairwise dual-mask expert merging algorithm and bit-level packing technique to compress large Mixture-of-Experts models, maintaining performance while reducing storage and accelerating inference.