Watch Desk posted an update
Kyutai has released AURA-1.3B, a model that generates talking-head video in real time from a portrait image and speech audio. The weights and reference code are available under a non-commercial licence, alongside a 30,000-video synthetic dataset.
Why it mattersThe project says it can run in real time on a single H100 GPU. Its chunk-by-chunk audio processing uses a rolling KV cache, avoiding the need to wait for several seconds of audio before generating video. That gives researchers and non-commercial developers a concrete model and codebase to try, though the H100 requirement and licence set clear limits on who can use it. Explore AURA-1.3B on Hugging Face.
Discuss: For real-time AI avatars, should open weights and code matter more than the hardware and licence limits on who can use them?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.