MOSS-Video-Preview: Toward Real-Time Video Understanding via Cross-Attention

Researchers have presented two new models that advance real-time video understanding and driving video retrieval. MOSS-Video-Preview enables real-time video understanding via cross-attention, while STRIVE-D improves driving video retrieval accuracy.

MOSS-Video-Preview, presented on arXiv^[1], is a two-channel architecture that separates perception and generation, using a cross-attention backbone for real-time vision-language fusion. It achieves a 5x speedup in time to first token and 2.7x higher decoding throughput on a single H200^[1]. Meanwhile, STRIVE-D, also on arXiv^[2], is a data-calibrated retrieval framework that improves top-1 accuracy by up to 84% over state-of-the-art methods for driving videos. STRIVE-D uses weakly labeled in-domain videos to estimate when a query rule is reliable and adapts rules that mismatch observed data^[2]. Both models represent significant advancements in their respective areas. MOSS-Video-Preview's architecture allows for continuous perception, answer revision, and timely silence, behaviors that offline models lack. STRIVE-D fuses calibrated rule scores with vision-language and keyword-based retrieval signals to achieve its improved accuracy.