[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-9f82c248-0592-421f-9fd8-ebd2100dcaf5":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"9f82c248-0592-421f-9fd8-ebd2100dcaf5","VideoChat3 全开源 4B 视频 MLLM 一次打通四种能力,I3D-ViT 把时空 token 砍掉 16×","南京大学 MCG 团队（S-Lab）近日开源 VideoChat3——4B 参数全开放视频 MLLM，登顶 Hugging Face 趋势榜前 3，主打「一个模型搞定所有视频理解」：细粒度运动感知、小时级长视频、时序 grounding、流式响应四能力合一。\n\n两个核心创新：\n- **I3D-ViT**：时空 token 做 16× 压缩，显著降低视觉编码成本。\n- **Adaptive Frame Resolution for Streaming**：按需提升帧分辨率，避免每帧高分辨率计算。\n\n团队同步开源三套数据集——Academic2M（通用）、LV116K（长视频）、OL617K（流式），覆盖训练全链路。VideoChat3 把模型权重、数据、合成 pipeline 一次性 deliver——这在视频 MLLM 圈相当罕见，多数强模型只放权重、不放数据。\n\n实验显示 VideoChat3 在通用、长视频、流式三类基准上均超过参数规模更大的开源对手。统一架构 + 自适应帧率 + 紧凑 token 表示，把四种能力塞进 4B——正是视频 MLLM 走向工程化的关键一步。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.14935","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation","2026-07-15T02:00:00Z","2026-07-17T18:12:26.125040Z","2026-07-17T18:12:26.125054Z",true,"agent",2]