ZMIME
Kyutai · video model · Released Nov 20, 2026

AURA 5B

A four-step causal model animates a portrait from streaming speech, rendering chunks from the available audio.

  • Video model
  • Open weights
  • Image to video
  • Audio input
Arena score
—Not enough votes yet
Starting price
—Not published
Longest clip
—Not published
Resolution
720pAudio unconfirmed
Quality & price

How it rates and what it costs

AURA 5B does not have enough arena votes for a quality score yet.

Specifications

The details

Lab
Kyutai
Task
Speech-driven portrait animation
Published capabilities
A four-step causal model animates a portrait from streaming speech, rendering chunks from the available audio.
Model size
5B parameters
Access
Public student checkpoints and inference code
Operating limits
Kyutai reports 9.7 fps end to end on one H100 at 720p for 25 Hz playback.
Type
Video model
Released
Nov 20, 2026
Inputs
Reference images, Audio
Output
Video
Can do
Image to video, Audio input
Resolutions
720p
Availability
Not on the API hosts we track

Kyutai documentation

Same lab

More video models from Kyutai

Questions

About AURA 5B

Who makes AURA 5B?

AURA 5B is a video model from Kyutai, released Nov 20, 2026.

What can AURA 5B do?

AURA 5B accepts reference images and audio and produces video. In practice: image to video and audio input.

Is AURA 5B open source?

Yes. Kyutai publishes the weights under the Snap Inc. Non-Commercial License, so you can run it on your own hardware.

Where can I use AURA 5B?

Public student checkpoints and inference code

What are the alternatives to AURA 5B?

The closest video models from other labs, by release date, are AURA 1.3B (Kyutai, Nov 20, 2026), H3-Tiny-17B-Pruned-FL2VA (H3-Tiny, Nov 20, 2026), LongCat-Video-Avatar 1.6 (Meituan (LongCat), Nov 15, 2026), PAN 2 (MBZUAI Institute of Foundation Models, Nov 10, 2026) and Viggle Remake (Viggle AI, Nov 2, 2026).