Wan Streamer 0.2

Wan Streamer v0.2 is a native-streaming audio-visual interaction model. User and agent text, audio and video share one causal timeline and are modeled by a single Transformer. It is an upgrade of v0.1 that raises the interactive output stream from 192×336 to 640×368 while keeping the ~200 ms model-side latency. The authors report about 550 ms total latency with 350 ms of network round trip.

For deployment, the model, which is trained end-to-end, is split into a thinker and a performer. A single-GPU thinker handles streaming perception, state update, K/V construction and decoding. A multi-GPU performer runs Ulysses-style context-parallel latent video generation. Work is pipelined across 160 ms streaming units. Compared with v0.1, v0.2 supports higher-fidelity close-up calls and scene-grounded mid-shot agents, where body and scene context are visible. Scenes start from a natural-language prompt and go live in about 0.5 s.

The page presents landscape and portrait demos, which it says are not sped up and include network time. Only a technical report (arXiv 2607.04443, submitted 5 Jul 2026) is referenced. No public weights, code, API or pricing appear in the provided material. Real-time operation needs multiple GPUs for the performer.