Wan Streamer v0.1: Single-Transformer Model Delivers Synchronized Audio-Video Responses Under 600ms

Wan Streamer is a new end-to-end multimodal foundation model built for real-time, full-duplex audio-visual interaction within a single Transformer architecture. Unlike existing systems that chain external ASR, language, TTS, and animation modules together, Wan Streamer processes and generates language, audio, and video as interleaved token sequences coordinated by block-causal attention — enabling streaming units as short as 160ms at 25fps.

The model achieves approximately 200ms of model-side response latency and around 550ms total end-to-end latency when accounting for network overhead, putting it under the 600ms mark and making it, according to its authors, the only system to deliver synchronized audio and video output from a single end-to-end model in under one second. Competing full-duplex systems such as GPT-4o Realtime and Doubao Voice respond quickly but produce no video output, while avatar renderers like LPM 1.0 and Hallo-Live only report rendering-stage latency, excluding their external language model dependencies.