Wan-Streamer v0.3 Reframes Video as World Plus Event Stream

A new paper from the Wan team lands with an unusually clean reframing of what a real-time interactive video model even is, and the framing may end up mattering more than any single release. In Wan-Streamer v0.3, the authors argue that a video is a world plus an event stream: the world is the persistent context, including environment, scene, subjects, ambient acoustic conditions and voice characteristics, and the event stream is everything that changes inside it, including subject behavior, speech, and other sounds.

That decomposition matters because it turns real-time video into what the paper calls a general-purpose pretraining task over large amounts of real video: given a world and incoming input, predict how the world moves, changes, and responds in real time. The claim is that the resulting competence can be specialized to a broad family of real-time downstream tasks, and the specific instantiation here is full-duplex audio-visual interaction, where the agent's own speech together with free-form behavior is treated as the event stream it produces.

The operating numbers are inherited from v0.2. The system runs 640x368 video at 25 FPS with a 160 ms streaming unit, roughly 200 ms of model-side response latency, and about 550 ms total interaction latency under a 350 ms bidirectional network budget. That is a tight loop for a video-plus-speech agent, though the resolution is modest against consumer avatar demos, and the end-to-end figure only holds if the network budget holds.

The honest caveat is that the arxiv page is a framing paper, not a benchmarks paper. It doesn't publish head-to-head comparisons, it doesn't detail what changed from v0.2 beyond preserving the operating point, and the claim that the multimodal process is vision-language-action-like is suggestive of an action channel without showing how expressive it is in practice. Take the specifics as reported.

The reason to still pay attention: if the world plus event stream view holds up in follow-on work, teams currently gluing together separate speech, avatar, and video pipes are looking at a design vocabulary that just got a lot cleaner, and a single-model target to chase.