huggingface.co web signal

Kyutai Releases AURA, an Open-Weights Streaming Avatar Model That Runs at 20.7 FPS on a Single GPU

Summary

Kyutai researchers released AURA on Nov 19, an audio-driven streaming avatar system with two sizes: AURA 1.3B generates 576p video at 20.7 fps on a single GPU, and AURA 5B runs 720p at 9.7 fps. The system uses chunk-by-chunk causal generation with a rolling KV cache, eliminating the multi-second audio look-ahead that bidirectional models require. Training is a two-stage distillation: causal student with sliding-window attention, then Distribution Matching Distillation with backward simulation. Weights, 30K synthetic training videos and code are on Hugging Face.