打通感知-理解-交互链路，全栈视频理解大模型VideoChat3开源了

TL;DR - VideoChat3 is an open-source 4B multimodal model that unifies short-, long-, and streaming-video understanding. Its native spatiotemporal encoding and adaptive visual resolution reduce token and compute costs while supporting timely responses to live video.

• I3D-ViT jointly models short frame sequences before temporal pooling, achieving roughly 16× spatiotemporal compression.
• A Silence/Standby/Response mechanism dynamically allocates visual resolution and lets the model decide when sufficient evidence exists to answer.
• The release includes model weights, code, and three datasets covering academic, long-form, and online video tasks.
• At 2,048 frames, reported latency fell from Qwen3-VL’s 44.449 seconds to 20.412 seconds, alongside lower FLOPs and memory use.