Alibaba reveals Wan Streamer for real-time video chat

Talking to an AI on a screen has always felt unnervingly disjointed. You speak, wait through an awkward pause, and watch a pre-rendered avatar flap its lips seconds after the audio starts playing.

Alibaba is trying to kill that lag entirely. On June 24, the company revealed Wan Streamer v0.1, an end-to-end foundation model built for sub-second, full-duplex video conversations.

Ditching the patchwork pipeline

Most existing video agent setups are awkward Frankenstein monsters. They chain together automatic speech recognition, a large language model for reasoning, a text-to-speech engine, and an external rendering module to animate the face. Every handoff between those systems adds lag and introduces sync errors.

Wan Streamer throws that entire pipeline out the window. Instead, it uses a single Transformer to handle text, audio, and video tokens all at once, coordinated by block-causal attention. It processes and generates speech and visuals natively inside the same model.

Because the entire stack is optimized for streaming, chunk sizes drop as low as 160 ms at 25 fps. Alibaba reports a model-side response latency of roughly 200 ms. Even after adding an estimated 350 ms for round-trip network transmission, total interaction latency hovers around 550 ms.

Just as important, the model is natively full-duplex. It does not wait for you to stop talking before it starts thinking. It watches and listens continuously while generating its own response, meaning you can interrupt it naturally just like you would with a human.

Natural gestures, but stuck at 192p for now

In released unedited video clips, the AI avatars look surprisingly alive. They casually brush hair out of their eyes, wave, and react with subtle facial expressions while chatting about daily fatigue or solving word riddles.

However, you cannot jump into a live call with it just yet. The v0.1 release is strictly a proof of concept, and the video stream currently renders at a tiny 192p resolution. Alibaba emphasized that this early version is meant to validate the end-to-end architecture, promising that scaling to higher resolutions will follow.

Why it matters

For everyday users, Wan Streamer is not something you can deploy today, and nobody wants to squint at a 192p video companion. But it gives us a clear look at where interactive AI is heading. The era of taking turns with a voice assistant is coming to an end, paving the way for visual agents that catch your micro-expressions and react before you even finish your sentence.

For the industry, Alibaba's approach poses a direct threat to the modular digital human ecosystem. Companies that built businesses around stitching together speech-to-text, LLMs, and 3D avatar rendering might find themselves obsolete if a single unified model can handle perception and generation in one pass.

We think betting on a unified Transformer for audio-visual streaming was the right architectural move. That said, the real test lies in the economics. Streaming 25 frames of video per second per user is computationally brutal. Pushing this model from 192p to crisp 1080p without bankrupting server infrastructure will be Alibaba's steepest hill to climb.