What Kyutai shipped

AURA takes a reference image and a stream of speech audio and generates a talking face as the audio arrives. That is a different job from the video models that render a finished clip. A conventional talking-head model reads several seconds of audio, runs dozens of denoising steps over the whole chunk, and only then shows a frame. Kyutai's TalkVerse teacher, the model AURA is distilled from, manages 0.58 frames per second on one H100.

AURA's answer is distillation with the direction of time fixed. The student generates one chunk at a time, attends only to the current chunk, the previous chunk and a fixed identity anchor from the start of the stream, and keeps a rolling cache so memory stays constant however long the conversation runs. Four denoising steps replace the teacher's fifty. The result is a 1.3 billion and a 5 billion parameter model, in a field where the paper notes streaming talking heads had been demonstrated only on much larger models of 14 to 18 billion parameters.

The release is unusually complete. Alongside both models' weights and the code, Kyutai published the 30,000 synthetic videos used for distillation, with the teachers' trajectories precomputed, so another lab can work from the same trajectories without first rebuilding the dataset. A live demo pairs the small model with Unmute, Kyutai's voice assistant, so a visitor can hold a spoken conversation with a generated face.

The three conditions on "real time"

The claim that travels is "real-time talking avatars on a single GPU." Kyutai's own pages support a narrower version of it.

First, the GPU. Every throughput number in the paper, including the comparison table against Hallo3, Wan-S2V and Live Avatar, is measured on a single H100, a data-center card. "Single GPU" is accurate and still worth saying out loud, because the gap between one H100 and a consumer card is where most readers' hardware sits. Nothing in the release shows AURA running at these speeds on a laptop or a gaming GPU. In the same table, on the same H100, Live Avatar runs at 1.9 fps, Hallo3 at 0.22 and Wan-S2V at 0.08.

Second, the resolution split. Only AURA 1.3B is real time, at 576p: it generates 20.7 frames per second for a stream that plays at 16. AURA 5B produces the sharper 720p picture at 9.7 frames per second, which is fast for a model of its size and still below playback rate. A product that needs 720p today does not get a live avatar from this release; it gets a near-live one.

Third, latency. Generating faster than playback is not the same as answering promptly. A chunk cannot be generated until its audio has been spoken, so Kyutai's streaming setup trails the voice by about two seconds with the paper model. For the public demo the team retrained the small model on half-second chunks and cut that to about 1.2 seconds. That is the pause built into every exchange in the demo setup, and it belongs in any honest description of the demo.

How far the quality claims go

On the authors' own tables, the small model is the surprise. AURA 1.3B records the best FVD (177.1) and identity-preservation score (CSIM 0.84) on the HALLO3 test set. The human study points the same way, within its limits: 27 raters and 400 pairwise comparisons preferred both AURA models over Hallo3 on all four axes, and rated the 5B model level with its own 50-step teacher on visual quality, lip-sync and naturalness, and higher on reference fidelity.

Two cautions sit inside those numbers. Lip-sync is the weak axis: AURA trails the larger Wan-S2V and Hallo3 on the automatic scores, and with human raters the 1.3B model ranks below its teacher on visual quality, lip-sync and naturalness, ahead only on reference fidelity. And length costs the bigger model. Across a minute of streaming, AURA 5B's identity score slides from 0.76 to 0.69 and its FVD worsens after the first ten seconds, then holds. The 1.3B model is steadier over the same minute, with a median per-video identity change of -0.02 and a few drifting outliers.

None of this has been replicated outside Kyutai yet. Treat the tables as a well-documented claim from the lab that built the model, with the unusual advantage that anyone with an H100 can test the throughput, the delay, the lip-sync and the identity stability on their own hardware.

What it opens up

The practical market is not film. It is the voice assistant that gets a face, the game character that answers in its own voice, the support agent a company runs on its own hardware instead of renting by the minute. AURA does not settle whether users want a generated face on those calls. What it changes is the test: the weights, the code and the trajectories are public, so the question can now be tried on one card instead of argued about.

Sources

  1. [1] Cohen, Dufour, Royer, Newson and Pérez, "AURA: AUdio-dRiven streaming Avatar", arXiv:2611.26540, submitted November 19, 2026Read source
  2. [2] Kyutai, "AURA: Audio-Driven Streaming Avatars in Real Time", project page, November 20, 2026Read source