✦
Summary
Wan Streamer v0.1 is a native-streaming, end-to-end interactive foundation model for real-time audio-visual interaction. It listens, sees, thinks, speaks, and generates synchronized video responses inside one Transformer, targeting full-duplex interaction rather than a pipeline of separate ASR, LLM, TTS, and avatar modules.
Key Features
End-to-end Transformer for synchronized speech and video responses.
Models language, audio, and video as both input and output tokens.
Uses block-causal attention for incremental streaming generation.
Targets full-duplex audio-visual interaction while continuing to perceive input.
Reports around 200 ms model-side response latency.
Supports 25 fps generation with short streaming units.
Avoids a multi-module ASR, LLM, TTS, and renderer pipeline.
Includes an arXiv paper and direct real-time recording video.
About Wan Streamer v0.1
The model represents language, audio, and video as interleaved input and output tokens coordinated by block-causal attention. Its stack uses causal encoders, causal decoders, low-latency multimodal scheduling, and streaming units as short as 160 ms to support around 25 fps interaction.
Wan Streamer is useful for real-time agents, interactive avatars, multimodal assistants, and research on low-latency full-duplex communication. The project reports roughly 200 ms model-side latency and about 550 ms total interaction latency when network latency is included.