Breaking

Alibaba's Wan Team Spotlights Wan-Streamer, a Real-Time Video Conversation Model, at ICML 2026

July 13, 2026 at 23:43 EDT

Media Generation
Foundation Models
Research & Papers

Alibaba's Wan team wrapped up its presence at ICML 2026 in South Korea, using a dedicated booth and a conference talk to showcase Wan-Streamer, an end-to-end model built for real-time, two-way video conversation.

ICML 2026 · Alibaba Wan Team

Wan-Streamer: One Model for a Real-Time, Two-Way Video Conversation

A single native-streaming Transformer collapses speech recognition, a language model, generation, and avatar rendering into one continuous sequence — enabling a live AI you can talk to face to face.

~550ms

End-to-end latency (≈200ms model-side)

25 fps

Real-time video output at 640×368

2×

A100-class GPUs assumed for the setup

Resolution jump: v0.1 → v0.2 (pixels per frame)

192×336

v0.1 · ~64k px

640×368

v0.2 · ~235k px

A ~3.6× jump in pixels per frame — while holding latency steady.

From pipeline to single model

Speech recognition → Language model → Generation → Avatar renderer = traditional chain

Thinker — low-latency reasoning → Performer — high-quality video = one native-streaming Transformer

Text, audio, and video handled as one continuous sequence via block-causal attention. No fixed avatar — users specify any character in natural language.

What impresses

Lifelike gaze, expressions, and gestures
One of the most human-feeling video calls seen
Small touches — thanking, saying it's "tired"

Where it stands

Research prototype — no public API or weights
Heavy video generation needs scaling work
Long-term stability and cost unresolved

Continue reading

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