Video = World + Event Stream

Overview概览

v0.3 broadens Wan Streamer from a conversation model into a general video learner. It separates what should stay coherent — the scene, characters, look, and sound — from what changes moment by moment, such as speech, motion, camera movement, and environmental events. We call these two parts the world and the event stream.v0.3 将 Wan Streamer 从对话模型拓展为更通用的视频学习模型。它把需要长期保持一致的场景、角色、画面与声音，和随时间发生的说话、动作、镜头移动、环境变化分开建模：前者是世界设定，后者是事件流。

That turns ordinary video into training material: establish the world once, follow the timeline, and learn what happens next. The resulting skill can be adapted to roaming, embodied control, or conversation; this release focuses on real-time audio-visual conversation.这样，普通视频就能直接用于训练：先确定世界设定，再沿时间线学习接下来会发生什么。得到的能力可迁移到漫游、具身控制或对话；本次发布聚焦实时音视频对话。

A robot navigates a suburban neighborhood and drives a car through a sequence of events — **Embodied navigation.** A fixed suburban world supports a causal event sequence: the robot runs through the neighborhood, enters a car, and begins driving. **具身导航。**固定的郊区世界设定支撑一条连续事件流：机器人穿过街区、进入汽车并开始驾驶。

A woman explains and demonstrates feeding chickens through a sequence of actions and speech — **Instructional activity.** Within a persistent poultry-farm world, speech, movement, and object interaction unfold together as time-aligned events. **讲解与操作。**在持续不变的养鸡场世界中，说话、移动和物体交互共同构成按时间对齐的事件流。

In a live session, the user's camera, microphone, and text drive the event stream. The model responds with speech and natural-language behavior cues, then renders both as synchronized audio and video.在实时会话中，用户的镜头、麦克风和文字共同驱动事件流。模型生成说话内容和自然语言行为指令，再把两者同步渲染为音频与视频。

Wan Streamer v0.3 native-streaming formulation with prefilled world context and parenthesized behavior — **How a live session runs.** The world is loaded once; user input, agent speech, behavior, audio, and video then advance on one shared timeline. **一次实时会话如何运行。**世界设定只需载入一次；此后用户输入、智能体说话与行为、音频和视频沿同一时间线持续推进。

Free-form Behavior自由行为

v0.1 and v0.2 learned to talk and visibly listen — with gaze, nods, micro-expressions, and lip sync. v0.3 adds actions written in ordinary language, interleaved with speech:v0.1 与 v0.2 主要学习说话和可见的倾听反应，例如视线、点头、微表情与口型同步。v0.3 进一步加入用自然语言描述的动作，并让它们与说话交织出现：

(reaches into the grass and picks up a green leaf) Look what I found.
(covers mouth in surprise) I did not expect that.（伸手到草丛中拾起一片绿叶）看看我发现了什么。
（捂住嘴，露出惊讶的表情）我没想到会这样。

The parentheses can hold anything the model can describe: pick up an object, turn toward a sound, change posture, or react. Because action and speech share the same stream, their timing is learned together and rendered directly — not choreographed afterward.括号里可以写任何可描述的行为：拿起物体、转向声音来源、改变姿态或做出反应。动作与说话共享同一条流，因此时机由模型一起学习并直接生成，而不是事后编排。

v0.1 / v0.2 / v0.3v0.1 / v0.2 / v0.3 对比

v0.2 made the stream clearer. v0.3 makes it more general and more expressive, while keeping v0.2's speed and resolution.v0.2 让实时画面更清晰；v0.3 在保持速度与分辨率的同时，让模型更通用、角色行为更丰富。

Aspect项目	v0.1	v0.2	v0.3
Main step核心进展	End-to-end live A/V端到端实时音视频	Higher resolution分辨率提升	General video learning + free behavior通用视频学习 + 自由行为
Training focus训练重点	Interaction data交互数据	Interaction data交互数据	World-event video pretraining, then interaction世界-事件视频预训练，再适配交互
Agent behavior角色行为	Speech + listening说话 + 倾听	Same, higher fidelity同上，更高保真	Speech + free-form actions说话 + 自由动作
Video视频	192×336 · 25 fps	640×368 · 25 fps	640×368 · 25 fps
Latency延迟	~200 ms model / ~550 ms total模型约 200 ms / 总计约 550 ms	~200 ms model / ~550 ms total模型约 200 ms / 总计约 550 ms	~200 ms model / ~550 ms total模型约 200 ms / 总计约 550 ms
Serving推理拓扑	Thinker + 1-GPU performerthinker + 单 GPU performer	Thinker + parallel performerthinker + 并行 performer	Same as v0.2与 v0.2 相同

Real-Time Full-Duplex Demos实时全双工交互演示

Live interaction results from the final full-duplex audio-visual model. The model listens, responds, and generates synchronized speech, video, and free-form behavior continuously in real time.由最终全双工音视频交互模型生成的实时交互结果。模型能够持续倾听和回应，并实时生成同步的语音、视频与自由行为。

NOT SPED UP. What you see is the actual end-to-end latency, including network time. No retiming or compositing. 未加速。 画面呈现的就是实际端到端延迟，包含网络耗时。无重定时或拼接。

World (condensed): A thoughtful young German engineer presents personal projects in a precise, well-organized home workshop with natural tool ambience. 世界设定（精简）：一位沉稳的年轻德国工程师在整洁专业的家庭工作室中介绍个人项目，伴有自然的工具环境声。

World (condensed): A warm, expressive young Italian chef shares family recipes in a golden-lit home kitchen with Mediterranean music and natural cooking sounds. 世界设定（精简）：一位热情亲切的年轻意大利厨师在暖金色家庭厨房中分享家传食谱，伴有地中海音乐与自然烹饪声。

Video Pretraining Demos视频预训练演示

Muted, offline results from the original, undistilled video pretrained model. These selected examples highlight action control across general scenes rather than real-time interaction.由未经蒸馏的原始视频预训练模型离线生成，所有样例均为静音。这些精选案例主要展示模型在通用场景中的 action 控制能力，而非实时交互。

v0.3 focuses on real-time full-duplex conversation. Extending generic action control to broader interactive scenarios is left for future work.v0.3 当前聚焦实时全双工对话；将通用 action 控制进一步扩展到更广泛的交互场景，留待未来工作。

Object handling. Move and place bowls, then pour the chocolate mixture. 物体操作。移动并摆放碗具，然后倒入巧克力混合物。

Cooking. Add paprika to the pan and stir the ingredients with a wooden spoon. 烹饪。向锅中撒入红椒粉，并用木勺搅拌食材。

Cleaning. Wash a pot, operate the faucet, and rinse a wooden cutting board. 清洗。清洗锅具、控制水龙头，并冲洗木制切菜板。

Cite this work引用本文

@misc{huang2026wanstreamerv03worldeventstream,
  title         = {Video = World + Event Stream},
  author        = {Lianghua Huang and Zhi-Fan Wu and Yupeng Shi and Wei Wang and
                   Mengyang Feng and Cheng Yu and Chen Liang and Junjie He and
                   Chen-Wei Xie and Yu Liu and Jingren Zhou and Ang Wang and
                   Bang Zhang and Baole Ai and Chongyang Zhong and Jinwei Qi and
                   Kai Zhu and Pandeng Li and Peng Zhang and Wenyuan Zhang and
                   Xinhua Cheng and Yitong Huang and Yun Zheng and Yuxiang Bao and Yuzheng Wang and
                   Zhiwei Lin and Zoubin Bi},
  year          = {2026},
  month         = jul,
  note          = {Wan-Streamer v0.3, technical report}
}