Breaking

Alibaba unveils Wan-Streamer, a real-time duplex interaction model

July 10, 2026 at 01:40 EDT

Foundation Models
Media Generation
Research & Papers

Alibaba's video-generation team Wan has unveiled Wan-Streamer, a full-duplex conversation model that handles audio, video and text in a single model, listening while responding with synchronized speech and video. The team published a v0.1 paper in June 2026 and a v0.2 paper in early July, with previews shown on the official site and a Hugging Face collection.

July 2026 · Wan (Alibaba)

Wan-Streamer: one model that listens, thinks, and talks back — with video

A full-duplex, omni-modal Transformer that understands a user's audio and video in real time and replies with synchronized speech and video — collapsing the old cascade of separate modules into a single native-streaming system.

~200ms

In-model signal-to-signal latency

25 FPS

v0.2 video frame rate

~160ms

Streaming unit size

v0.1 → v0.2

Resolution jumps, latency holds steady

~64K px

v0.1

192×336

→

~235K px (≈3.7×)

v0.2

640×368

K/V cache sharing + parallelized denoising on a multi-GPU Performer raised the pixel count ~3.7× — while in-model latency stayed at ~200ms.

The ~550ms round trip

~350ms

Network budget

~200ms

In-model

~550ms

Total remote

Inside the architecture

Split brain: Thinker + Performer on a shared causal timeline

Thinker · single GPU

Low-latency perception, reasoning and decoding

→

Performer · multi-GPU

High-resolution, synchronized video generation

Block-causal attention interleaves input and output text, audio and video — and characters aren't fixed avatars: any human, pet or anime style can be described in natural language.

What sets it apart

Native integration of video output in a full-duplex model — going beyond voice-only systems like GPT-Live, and replacing brittle cascaded pipelines prone to latency and error build-up.

Still unproven

An early research preview: generation quality, speed vs. rivals, stability (e.g. game NPCs), hallucination and long-conversation limits all need verification. Pricing, release date and open-weights plans undisclosed.

Continue reading

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