AURA 1.3B vs H3-Tiny-17B-Pruned-Ref2VA
Newest: AURA 1.3B · Native audio: H3-Tiny-17B-Pruned-Ref2VA. Not enough arena votes yet for a quality score, so this is a spec comparison.
Kyutai
AURA 1.3B
- Arena—
- From—
- Max length—
H3-Tiny
H3-Tiny-17B-Pruned-Ref2VA
- Arena—
- From—
- Max length—
Add a model
Make it a three-way comparison.
Where they differ
Quality is the arena score from blind, side-by-side human votes. Price, length and resolution come from each lab’s own listings.
- NewestAURA 1.3BAURA 1.3B Nov 20, 2026 · H3-Tiny-17B-Pruned-Ref2VA Nov 19, 2026
- Widest inputsSame inputsAURA 1.3B: Reference images, Audio · H3-Tiny-17B-Pruned-Ref2VA: Text prompts, Reference images
- Native audioH3-Tiny-17B-Pruned-Ref2VAGenerates an audio track with the video: AURA 1.3B unconfirmed · H3-Tiny-17B-Pruned-Ref2VA yes
- Image to videoBoth animate imagesStarts from an uploaded image: AURA 1.3B yes · H3-Tiny-17B-Pruned-Ref2VA yes
- Self-hostingAll open weightsDownloadable weights (CC BY-NC 4.0 and MiniMax H3 Community License)
- Most API hostsSame reachAURA 1.3B: 0 · H3-Tiny-17B-Pruned-Ref2VA: 0
Every spec in one table
Dashes mean the data is not published.
| Specification | ||
|---|---|---|
| Task | Speech-driven portrait animation | Audiovisual generation |
| Published capabilities | A four-step causal model animates a portrait from streaming speech, rendering chunks from the available audio. | Generates video and synchronized sound from reference imagery using a pruned MiniMax H3 transformer. |
| Model size | 1.3B parameters | 17.07B transformer parameters |
| Access | Public student checkpoints and inference code | Downloadable transformer checkpoint for the H3 pipeline |
| Operating limits | Kyutai reports 20.7 fps end to end on one H100 at 576p for 16 Hz playback. | 50 steps |
| Sources | Model weights | Model weights |
| Output | ||
| Produces | Video | Video with sound |
| Resolutions | 576p | — |
| Highest resolution | — | |
| Sound | — | Always generated |
| Inputs | ||
| Text prompts | No | Yes |
| Start image (image to video) | Yes | Yes |
| Video input | No | No |
| Audio input | Yes | No |
| Access | ||
| Weights | OpenCC BY-NC 4.0 | OpenMiniMax H3 Community License |
| Released | Nov 20, 2026 | Nov 19, 2026 |
Which should you choose?
Which is better, AURA 1.3B or H3-Tiny-17B-Pruned-Ref2VA?
Neither model has enough arena votes for a reliable quality score yet, so this page compares specs only. Run the same three or four prompts through each and judge the results.
What is the difference between AURA 1.3B and H3-Tiny-17B-Pruned-Ref2VA?
AURA 1.3B is a video model from Kyutai (Nov 20, 2026), taking reference images and audio and producing video. H3-Tiny-17B-Pruned-Ref2VA is a video model from H3-Tiny (Nov 19, 2026), taking text prompts and reference images and producing video with sound.
Which is newer?
AURA 1.3B Nov 20, 2026; H3-Tiny-17B-Pruned-Ref2VA Nov 19, 2026.
Which generates video with sound?
Generates an audio track with the video: AURA 1.3B unconfirmed; H3-Tiny-17B-Pruned-Ref2VA yes.
Can I run any of these on my own hardware?
All open weights. Downloadable weights (CC BY-NC 4.0 and MiniMax H3 Community License).