Breaking

Alibaba unveils Wan-Streamer, a single model generating synced speech and video

July 10, 2026 at 02:13 EDT

Foundation Models
Media Generation
Research & Papers

The Wan team led by Alibaba's Tongyi Lab has introduced Wan-Streamer, a duplex interaction model that listens, watches, understands and speaks through a single Transformer, generating synchronized speech and video at once. The latest v0.2 was posted to arXiv in July 2026, with a paper released alongside a dedicated site, wan-streamer.com. Its predecessor v0.1 appeared on arXiv in June the same year, positioning the work as a continuing research line.

July 2026 · Tongyi Lab

Wan-Streamer: one Transformer that listens, watches, understands and speaks

A duplex, end-to-end omni-modal model that generates synchronized speech and video at once — collapsing the usual VAD → ASR → LLM → TTS → animation pipeline into a single native-streaming network.

~200ms

Model-side signal-to-signal latency

~550ms

Total interaction latency (incl. 350ms network)

160ms

Streaming processing unit

25fps

Output frame rate (v0.2)

Output resolution: v0.1 → v0.2

Pixel area grows from proof-of-concept close-up to mid-shot scenes.

v0.1 · June 2026

192×336

close-up

→

v0.2 · July 2026

640×368

mid-shot scenes

≈ 3.6× more pixels — plus clearer expressions, gaze, gestures and scene context.

From cascaded pipeline to a single model

VAD → ASR → LLM → TTS → Animation → Video gen

Old way: latency accumulates and modalities drift out of sync.

Thinker

Low-latency perception, reasoning & decoding on a single GPU.

Performer

Heavy video generation across multiple GPUs via Ulysses-style context parallelism.

One native-streaming Transformer with block-causal attention jointly learns perception, reasoning, generation, turn management and cross-modal sync — no fixed avatar rig.

What draws praise

Natural real-time video conversation that watches, listens and speaks
Native single-model streaming extending all the way to video output
Talks as any character describable in words — human, pet, anime

Open questions

v0.1's ~192p resolution seen as proof-of-concept
Whether open weights will be released is undetermined
Production stability and diverse use-case validation still future work

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.