Alibaba's Wan Team put Wan-Streamer v0.2 (arXiv:2607.04443) online — an engineering upgrade of "latency unchanged, resolution multiplied several times". v0.1 had already packed full-duplex audio-video interaction into a single Transformer's unified causal timeline, at the cost of only 192×336 output, enough for video-call close-up but in mid-shot, character posture, desk objects, and surrounding environment all blur into a mess. v0.2's goal: lift resolution from 192×336 to 640×368 (about 3.5× more pixels), frame rate still 25 FPS, model-side signal-to-signal latency stay at ~200ms, total remote interaction latency with the 350ms two-way network budget held at ~550ms. This means the upgrade can't touch that latency-sensitive causal path; new compute can only be shunted to non-latency-critical links. The solution is a re-partitioning of the Thinker-Performer deployment topology. Thinker continues to reside on a single card, handling streaming perception, short language/state Transformer, KV-cache construction, and the final decoding; Performer becomes a Ulysses-style context-parallel multi-card group, specifically handling long-sequence latent-space denoising: each rank maintains a Ulysses-sharded local KV cache, the high-resolution latent video sequence does all-to-all/gather between ranks, and short audio latents aren't sharded. Thinker only broadcasts the KV slice that the performer can consume, the language state itself doesn't need to sync across cards, and the remote-end latency stays around ~550ms. On the visual side, v0.2 makes close-up calls clearer, and for the first time supports "in-scene mid-shot digital human": sitting posture, eye direction, hand movement, and desk items remain readable in real-time conversation, and the digital human is no longer locked in the face-framing box. This is a step pushing full-duplex interaction from "dialogue" to "dialogue in a scene", with direct impact on customer-service seats, virtual streamers, and companion robots for the final viewing experience. From a larger perspective, this path shows that real-time audio-video generation is no longer "inference-speed optimization" but has entered the "streaming-causal + deployment-topology co-design" phase — next-unit streaming modeling, context-parallel performer, low-latency thinker guarding, the three have formed a complete engineering template; real-time voice+vision systems like Grok and GPT-4o will eventually take the same topology-tiering route.