PARAS MADAN
All writing

Latest in AI

Kyutai AURA 1.3B vs 5B: How to Test a Live Talking Avatar

Compare AURA’s live playback and offline rendering paths, then measure first-frame delay and lip sync before deciding which checkpoint fits.

Kyutai released AURA on November 20. It turns a portrait and speech into video as the audio arrives. If you want a face for a live voice prototype, start with 1.3B. The 5B model produces a higher-resolution image, but Kyutai’s own measurements show it does not generate frames fast enough for its intended playback rate on one H100.

The spec I would check first is the gap between generation and playback. That tells you whether the avatar can keep talking without the video falling progressively behind the audio. It still does not say how long the viewer waits for the first frame.

Choose the checkpoint by playback rate

Kyutai’s reference repository separates headline measurements from the speed of its released code. Both sets of numbers below are for one H100, and neither is a promise for another GPU.

CheckpointOutput and target playbackOptimized research pipelineReleased code with compilation
AURA 1.3B576p at 16 fps20.7 fpsAbout 16 fps
AURA 5B720p at 25 fps9.7 fpsAbout 8 fps

The 1.3B reference result is roughly at its playback target, so there is little evidence here of spare capacity for simultaneous streams. At 9.7 generated frames a second, 5B would take roughly 77 seconds to produce the 750 frames needed for 30 seconds of 25 fps playback, assuming a steady rate and ignoring startup. That makes 5B a reasonable offline comparison, not the baseline for a live conversation.

Use the public demo to see the workflow

Kyutai’s live demo pairs AURA 1.3B with Unmute. The demo answers a simple question: does it feel useful to put a moving face on the voice? It does not isolate the avatar’s delay from the voice system, network and playback buffer.

For a controlled check, use a synthetic or consented portrait and a short recording you have permission to use. The repository documents a setup with a CUDA GPU and uv, plus a file-rendering example that takes a portrait and a WAV and writes an MP4. Start there to see whether the face stays recognisable and the mouth movements line up. Finishing an MP4 successfully does not establish live performance.

Measure first frame, playback and drift separately

AURA works chunk by chunk without waiting for future audio. It does still need the audio for the current chunk. Time the delay from the first audio arriving at your avatar service to the first visible frame. If you are testing a synthetic voice system as well, measure from the start of the entire interaction in a separate column.

Then play a longer clip containing quick speech, pauses and a few hard consonants. Log the audio and video timestamps at the beginning and end. If the gap widens, the renderer is behind even if the first few seconds looked fine. Repeat after startup and after the renderer is warm, and then repeat with two streams before estimating GPU capacity.

One worthwhile control is the same voice with a static portrait. If the avatar makes the response slower without making the conversation easier to follow, you learn something useful before spending time on a higher-resolution checkpoint.

Read the weight licenses before a product pilot

Kyutai lists CC BY-NC 4.0 for the 1.3B weights and Snap’s non-commercial terms for the 5B weights. Access to the code and checkpoints does not give you permission to build a commercial avatar product with those weights. Check the license for the artifact you plan to use, and keep the portrait and audio permissions separate.

For a research prototype, I would start with one face, one voice and one stream on 1.3B. A stable picture that follows the audio is more useful than a larger frame that arrives late.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain model releases and technical trade-offs in writing that developers can use.

Book a 30-minute call.