Research Debrief · XiaoHu Explains

Wan Streamer: Real-Time AI That Watches, Listens, and Talks

Model-side response is about 200ms, total latency about 550ms; v0.1 is only 192p, and the demos are pre-recorded, not live real-time experiences.
At a Glance
  • Wan Streamer v0.1 is a natively streaming, end-to-end interactive foundation model that models language, audio, and video as both input and output within a single Transformer, coordinated by block-causal attention, generating in a streaming fashion as data arrives.
  • The model itself takes about 200ms to compute a response; add 350ms of round-trip network latency and total interaction latency comes to about 550ms. At 25fps, the shortest streaming processing unit is just 160ms.
  • The team claims it's the only model that can output synchronized audio and video from a single end-to-end Transformer with total latency under 1 second; existing systems either output audio only (GPT-4o Realtime, Doubao, Gemini Live) or are stitched together from a chain of modules: ASR + LLM + TTS + animation.
  • The current v0.1 resolution is only 192p, positioned as a proof of concept for the end-to-end design; the character demos released are unedited, pre-recorded model outputs, not live real-time experiences.
  • The latency numbers in the comparison chart mix different measurement scopes: the top group covers the full end-to-end interaction loop, while the bottom group only counts the rendering stage (excluding external LLM/ASR/TTS); the team notes these should be read with that caveat in mind.
⚑
This is Wan Streamer's official launch page (a vendor release). The latency figures, capability comparisons, and positioning language like "only" all come from the vendor; some comparison numbers mix different measurement boundaries, which is noted inline where relevant.
1What This Is

One Model Handles Real-Time Audio-Video Conversation

The Tongyi Wanxiang (Wan) team recently released Wan Streamer v0.1, a real-time audio-video interaction model. There's no shortage of AI that can talk with you in real time these days, but almost none can watch your face, listen to you speak, respond out loud, and animate its own moving face all at once. Wan Streamer packs all of that into a single model.

It handles language, audio, and video input and output all within a single Transformer, achieving sub-second full-duplex audio-video conversation: the model itself takes only about 200 milliseconds to compute a response, and with network round-trip added, total latency comes to about 550 milliseconds.
⚡

Why it's worth a look: real-time conversational systems today fall into two camps — one responds fast but only produces audio, with no visible face (GPT-4o Realtime, Doubao, Gemini Live); the other has a face, but it's stitched together from external ASR, a language model, TTS, and animation modules. The team calls Wan Streamer the only model that outputs synchronized audio and video from a single end-to-end Transformer with total latency under 1 second.

~200 ms
Model-side response latency (as reported by the team)
~550 ms
Total interaction latency = 200ms model-side + 350ms network
160 ms
Shortest streaming processing unit at 25fps
192p
v0.1 resolution, a proof of concept for the end-to-end design
Breaking down the 550ms total latency
Model-side 200ms
Round-trip network 350ms
The model itself accounts for only 200ms; the remaining 350ms is network round-trip. In other words, the model's raw reaction speed is faster than the headline total latency suggests.
2See It in Action

Four Character Demos + a Real-Time Screen Recording

The four demos below all come from the same model — only the character, voice, and setting change. Let's look at the results first, then get into how it works.

Before you watch: these are all unedited, pre-recorded model outputs, not live real-time experiences; the current v0.1 resolution is only 192p, used to validate whether the end-to-end design works. The team says scaling to higher resolution later should be relatively easy — but that's a plan, not something v0.1 has already achieved.