Wan Streamer v0.3

Wan Streamer v0.3 is a research model from Alibaba that learns video as a persistent "world" (scene, characters, look, sound) plus an "event stream" (speech, motion, camera movement, environmental changes). This lets ordinary video be used as training material, and the learned skill can in principle be adapted to roaming, embodied control or conversation. The v0.3 release focuses on real-time full-duplex audio-visual conversation.

In a live session, the user's camera, microphone and text drive the event stream. The model replies with speech plus natural-language behavior cues in parentheses, such as picking up an object or turning toward a sound. It renders both as synchronized audio and video. Earlier versions (v0.1, v0.2) covered speaking and visible listening, such as gaze, nods and lip sync. v0.3 adds free-form actions interleaved with speech.

The project page lists output of 640×368 at 25 fps, with about 200 ms model latency and about 550 ms total latency (stated for v0.1/v0.2 and the same speed as v0.2). Serving uses a thinker plus a parallel performer. The page notes that extending general action control to broader interactive scenarios is future work. According to a YouTube creator summary, only a paper has been released. No code, weights or public product were found in the provided material.