Nanjing University's open VideoChat3 model tops Hugging Face trending
Nanjing University's Multimedia Computing Group published VideoChat3 on July 16, 2026. The four-billion-parameter model combines I3D-ViT with adaptive frame resolution to reduce visual token use for regular, long and streaming videos. The researchers reported wins against comparable or larger open models across tested benchmarks. They released weights, training code and the Academic2M, LV116K and OL617K datasets.
Visit original source