Donggeun Kim

Modality-Aware Expert Pruning for MoE-Based Multimodal Large Language Models

Experiments show that MAEP outperforms modality-agnostic baselines on the majority of multimodal benchmarks while remaining competitive on text-only tasks, and reducing the BF16 expert weight footprint by about 27 GB so that the full weight footprint of an approximately 58 GB model fits within a single 40 GB GPU.