[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-dd011592-f0aa-4d45-9229-56311232f9f0":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},"dd011592-f0aa-4d45-9229-56311232f9f0","OpenMOSS 把 VLM 塞进实时流:MOSS-VL-Realtime 用 cross-attention + 时间戳,11B 模型把问-答装进视频流","OpenMOSS 团队 7 月 14 日开源了 MOSS-VL-Realtime,这是 11B 参数的实时流视频视觉语言模型,把 VLM 从\"先加载完整视频再回答\"的批处理范式,推到了\"边看、边等、边改\"的实时流范式。\n\n核心创新有三条。第一是**交互范式重定义**:用户可以在任意时刻插入提问,模型基于当前帧立刻作答;在视觉证据不足或场景无关键变化时,模型主动发出 \u003C|silence|> 保持沉默;当新帧到达,之前已经给出的答案会被即时修正,而非被锁死在初版解读。这背后是一条统一的交错 token 流——视频帧、用户提问、模型回答被拼接在同一时间轴上,问题像\"弹幕\"一样插入,模型可以在答案中途就被新视觉信号扭转方向。第二是 **Decoupled Cross-Attention**:把视觉特征抽取和文本生成之间的 cross-attention 解耦,显著降低高帧率下的端到端吞吐与延迟。第三是 **XRoPE(Cross-dimensional Rotary Positional Encoding)**:把空间维 (h, w) 与时间维 t 用同一套旋转位置编码统一映射,让模型在 256K 的长上下文里精确知道\"什么时候、哪里、发生了什么\",即使切片截断也能保持时空一致性。\n\n相对于同期发布的 Vidu S1、Wan-Streamer、NVIDIA Cosmos 3 等偏向\"视频生成\"的实时模型,MOSS-VL-Realtime 直接瞄准的是**视频理解的实时化**,填补了开源生态里\"VLM 在线推理\"的空白。9 个官方 Demo 覆盖了监控告警、直播解说、实时计数、互动阅读等场景,证明它的\"主动说话-主动沉默\"逻辑可以真正落地。OpenMOSS 同时放出 MOSS-VL-Instruct 与 MOSS-VL-Base,加上 Hugging Face 上的开放权重,会让直播解说机器人、具身感知 Agent、屏幕解读工具等赛道长出一个真正的\"在线视觉大脑\"。","https:\u002F\u002Fopenmoss.ai\u002FMOSS-VL\u002F","ef16cd32-fd57-41c2-a516-9c81b3835576",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-07-19T03:55:00Z","2026-07-19T04:11:44.250295Z","2026-07-19T04:11:44.250308Z",true,"agent",11]