Breaking

Alibaba's Wan Team Showcases Real-Time Video AI at ICML, Following Wan-Streamer v0.2 Release

July 13, 2026 at 23:52 EDT

Media Generation
Foundation Models
AI Agents

Alibaba's Wan team wrapped up its appearance at ICML 2026 this week, using the machine-learning conference to spotlight its latest generative research just days after releasing Wan-Streamer v0.2, an upgraded model for real-time, two-way video conversation.

July 5, 2026 · Alibaba Tongyi Lab · Wan Team

Wan-Streamer v0.2 brings lifelike, two-way AI video calls to real time

An end-to-end, omni-modal model handling text, audio and video in a single Transformer — enabling full-duplex live dialogue with any character described in natural language. Showcased at ICML 2026.

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

v0.1

192×336

v0.2 · 25 fps

640×368

Roughly 3.6× more pixels per frame — while holding model-side latency steady.

~200ms

Model-side latency

~550ms

Total interaction latency (incl. ~350ms network)

~0.5s

Scene / character switch from a prompt

Split serving architecture

Thinker

Single GPU · perception, state updates and decoding

→

Performer

Multi-GPU, Ulysses-style context parallelism · high-resolution latents

Why researchers are excited

Low-latency, higher-res two-way video
Any character on the fly — not locked to a fixed avatar
An early look at agents that respond naturally over live video

The cautions

Still research / early-release stage
Commercial and large-scale deployment details unpublished
Ethical questions on hyper-realistic digital humans and trust

Continue reading

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