Event

Nanjing University and others release open-source VideoJev video understanding model

1 reports · 1 sources

Nanjing University, JD.com and other institutions have open-sourced the VideoJev video understanding model, providing weights in two sizes, VideoJev-4B and VideoJev-35B-A3B (approximately 35B total parameters, approximately 3B activated parameters), along with core training, evaluation, and inference code; the weights are released in BF16 and Safetensors formats. The model unifies video temporal grounding and selective video question answering as a candidate scoring task, and its input uses only video frames, without relying on subtitles or audio. On LVBench, the model achieves a score of 73.4%; according to its project description, on an H200 with batch size=1, each question takes about 1.5 seconds. This time includes visual encoding, prefix prefill, and model forward computation, is amortized over the number of questions in the full LVBench test set, and excludes preprocessing and data transfer.