OmniTide: Co-Designing Algorithms and Systems for Efficient On-Device Omni-LLM Streaming.

arXiv:2609.34653v2 Announce Type: replace Abstract: On-device streaming omni-modal inference safeguards user privacy and eliminates prohibitive per-token API costs, but faces a critical bottleneck: the continuous influx of multimodal data rapidly exhausts cons