Kyutai open-sources AURA: low-latency audio-driven streaming avatar via rolling KV cache

kyutai · hf · 2026-11-20

Kyutai released AURA (AUdio-dRiven streaming Avatar), a causal video-generation model for streaming applications. It combines a rolling KV cache with a two-stage distillation pipeline to drastically cut diffusion steps. The paper (with ISIR/Sorbonne and Google DeepMind collaborators) shows 5B and 1.3B variants, the latter running in real time; weights are on Hugging Face.

Original post →

More from Multimodal

Multimodal channel →