BREAKING
Alibaba unveils Wan-Streamer
Audio, video and text in one model
One model replaces the cascade
1
Perception
↓
2
Reasoning
↓
3
Generation
↓
4
Sync
Thinker and Performer split
Thinker
single GPU
●
Perception
●
Reasoning
●
Decoding
Performer
multi-GPU
●
High-res video
●
K/V cache sharing
●
Parallel denoising
0
px
width v0.2
0
FPS
frame rate
0
ms
total latency
Still an early research preview
AI NEWS BLITZ
Alibaba's Wan team has revealed Wan-Streamer, a real-time duplex interaction model.