VideoChat3 Ships a Fully Open Video MLLM Aimed at Generalist Video Understanding

arXiv / HuggingFace Daily Papers·low signal

Nanjing University's Multimedia Computing Group released VideoChat3 (arXiv 2607.14935, 124 upvotes on July 17's Daily Papers), positioned as fully open — weights and training recipe — for efficient generalist video understanding rather than a narrow captioning or QA head. The open-video-MLLM lane has been notably thinner than open text models, so a complete release here fills a real gap for builders who need video comprehension without an API dependency.