OmniPack: Training-Free Token Compression for Omni-modal LLMs
August 3, 2026
OmniPack is a training-free framework that optimizes Omni-modal LLMs by coordinating structural compression before the LLM with semantic refinement within the LLM. This approach reduces computational overhead from redundant audio-visual tokens while preserving essential structural evidence.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference costs and latency for multimodal models without requiring expensive retraining.