VideoChat3: Open-Source 4B Video Model Beats GPT-5 on Temporal Grounding, Halves Token Count
AI video model releases
VideoChat3: Open-Source 4B Video Model Beats GPT-5 on Temporal Grounding, Halves Token Count
VideoChat3-4B, from Nanjing University, Shanghai AI Lab, NTU, and Peking University, outscores GPT-5 and Gemini 2.5 Flash on all three TimeLens temporal grounding benchmarks: Charades-STA (56.1 vs 40.5 vs 48.6 mIoU), ActivityNet Captions (54.6 vs 42.9 vs 52.5), and QVHighlights (67.0 vs 52.1 vs 64.3).
The model generates exactly half the visual tokens of Qwen3-VL under identical inputs, producing 12,544 tokens at 256 input frames versus Qwen3's higher count.
The token savings come from the I3D-ViT (Inflated 3D Vision Transformer), which groups four consecutive frames per chunk and applies spatiotemporal self-attention before pooling, combined with 2x2 spatial downsampling for a 16x compression ratio.
The release includes model weights, training code, training strategy, and three custom datasets totaling about 3 million instruction samples, all published on the MCG-NJU Hugging Face collection on July 16, 2026.
The benchmark numbers are self-reported by the paper's 27 authors and have not been independently replicated yet; the paper reached #3 on Hugging Face's Paper of the Day on July 17.
cho.sh