Tasks and inputs
AURA-1.3B is listed for Image-to-video, Multimodal, Audio-Driven Video. It accepts image, audio, text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
AURA-1.3B is Kyutai's real-time streaming talking-head model: a portrait plus live speech becomes 576p, 16 fps video, chunk by chunk, on one GPU.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights, non-commercial; no hosted API price.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
A voice agent that can be seen, not just heard, needs a face that keeps pace with the audio. AURA-1.3B is Kyutai's answer, released on 18 November 2026: give it one portrait and a speech stream and it emits talking-head video in small chunks, each driven only by its own slice of audio. No multi-second look-ahead, no waiting for the sentence to finish.
Of the two AURA models, this is the one that streams in real time. Kyutai already pairs it with its Unmute cascaded voice system in a live demo.
| Property | Value |
|---|---|
| Output | 576×576 or 768×512, 16 fps |
| Steps per chunk | 4 (timesteps 1000, 745, 498, 255), no guidance |
| Speed, optimised pipeline | 20.7 fps end-to-end on one H100 (DiT 28.8 fps, VAE 96.9 fps) |
Speed, reference code with --compile |
about 16 fps on one H100 |
| Audio features | Kyutai's Mimi codec |
| Main checkpoint | stage2_dmd/step-6500_ema_merged.safetensors, ~6.3 GB fp32 |
| Licence | CC BY-NC 4.0 |
Both speed figures clear 16 fps playback, which is what "real time" means here. They are not the same number, though. The 20.7 fps comes from an optimised pipeline reported in the paper; the public reference code reaches roughly 16. If you need headroom for networking and audio, benchmark the code you will actually ship.
Memory stays flat because the rolling KV cache keeps just two things: the first chunk, as an identity anchor, and the most recent chunk. That is what lets a stream run indefinitely without the face drifting.
The model itself is a four-step causal student. Its teacher, AURA-1.3B-Teacher-TalkVerse, is a bidirectional, multi-step model built on Wan2.1-T2V-1.3B weights and conditioned on Mimi audio tokens. Distillation ran in two stages: an ODE regression onto 30,000 teacher trajectories (released as a dataset), then distribution matching distillation with the audio adapter frozen, using EMA weights for the final release. The teacher's weights are public for distillation research, but the inference code does not run it.
Clone the AURA repository, run uv sync, then uv run generate.py --image portrait.jpg --audio speech.wav --output out.mp4. The 1.3B model is the default. Missing pieces download automatically: the merged checkpoint, the Wan2.1 VAE and T5 text encoder (Apache-2.0) and Mimi. Square or landscape portraits where the face fills much of the frame work best.
On Kyutai's HALLO3 test split the model reports FID 56.8, FVD 177.1 and Sync-C 4.07, and the best identity preservation (CSIM 0.84) of every method compared. Lip-sync is good rather than class-leading. Run your own checks on three things: identity hold after a minute of speech, lip timing on fast consonants, and what happens with non-English audio, which Kyutai has not tested.
Licensing and ethics both limit deployment. The weights are non-commercial. Kyutai also asks that you animate only consenting subjects and label output as AI-generated.
Want sharper 720p and can live without real time? Look at AURA-5B, the larger sibling. Kyutai's other releases are on its provider page, and alternatives such as Wan Animate 2 are gathered under audio-driven video.
Compare the model on the work you need to complete.
AURA-1.3B is listed for Image-to-video, Multimodal, Audio-Driven Video. It accepts image, audio, text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.
Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.
The weights are downloadable under CC-BY-NC-4.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
kyutai
AURA-5B is Kyutai's larger streaming talking-head model: portrait plus speech to 720p, 25 fps video in four steps per chunk, at about 9.7 fps on one H100.
Alibaba
Wan-Dancer-14B is Alibaba Wan's open music-to-dance video model, planning a whole track before refining it into minute-scale coherent dance.
skywork
SkyReels-V3-A2V-19B is Skywork's open 720P talking-avatar model that animates a reference image from up to 200 seconds of driving audio.
Meituan LongCat
LongCat Video Avatar 1.5 generates audio-driven character video from speech, scene instructions and optional image references.