Open-source VideoChat3 beats GPT-5 across TimeLens temporal grounding tasks
A team from Nanjing University, Shanghai AI Laboratory, NTU and Peking University released VideoChat3 on July 16, 2026. Its four-billion-parameter video model outperformed GPT-5 and Gemini 2.5 Flash across TimeLens temporal grounding tasks. It supports motion, long videos and live streams. The MCG-NJU team released weights, code and training methods on Hugging Face, including three datasets containing roughly three million instruction examples.
Visit original source