TL;DR of the paper, generated automatically.

Training-free pruning that shrinks MoE multimodal LLMs by scoring experts per modality.

• Memory: cuts about 27 GB of expert weights, roughly 47% of total memory.
• Fit: the paper says a ~58 GB model then fits on one 40 GB GPU.
• Results: beats modality-blind baselines on most multimodal tests, competitive on text.