Alibaba's OmniPack Cuts Qwen2.5-Omni-7B FLOPs to 16.7% Without Retraining, Keeps 98% of Baseline Score

OmniPack, from Alibaba, PKU, Tsinghua and Northwestern Polytechnical, is a training-free token-compression framework for omni-modal LLMs that runs structural compression before the LLM and query-conditioned semantic refinement inside it. On Qwen2.5-Omni-7B it preserves 98.0% of the original score with just 16.7% of the FLOPs, and holds 92.9% at 6.8% of the FLOPs, outperforming existing compression methods across five audio-visual benchmarks and three backbones.