Tasks and inputs
AURA-5B 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-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.
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
AURA-5B trades real time for resolution. It is the bigger sibling in Kyutai's AURA release of 18 November 2026, producing 720p talking-head video at 25 fps from one portrait and a speech track, and it does so in a streaming, chunk-by-chunk way. What it cannot do on a single H100 is keep up with its own playback rate.
Kyutai's own description contains that tension. Its headline promises generation "in real time"; its list of limitations says plainly that 9.7 fps end-to-end sits below 25 Hz playback. The second statement is the accurate one. If a live avatar is the goal, Kyutai points to AURA-1.3B instead.
--compile: about 8 fps on one H100.The VAE is nearly as slow as the transformer here, so optimising one without the other buys little. For offline rendering, a podcast clip or a batch of explainer videos, 8–10 fps is perfectly workable.
Kyutai did not train this one from its own foundation. The teacher is Snap's 50-step bidirectional TalkVerse-S2V-5B, itself built on Wan2.2-TI2V-5B, and the student keeps that lineage. Three changes make it causal: block-wise causal attention over four-latent-frame chunks (about 0.64 s each), no motion-frame conditioning, and mixed diffusion timesteps per chunk. Audio comes from a wav2vec2 large XLSR-53 English encoder, not Kyutai's Mimi codec.
Training ran on 64 H100s in two LoRA stages (rank 128): ODE distillation, then distribution matching distillation. The final file, stage2_dmd/step-5000_merged.safetensors, is a standalone fp32 transformer of 5.7B parameters and 22.8 GB. LoRA-only and stage-1 checkpoints are there for researchers.
Because every file derives from TalkVerse, the weights inherit the Snap Inc. Non-Commercial License. The Wan2.2 VAE and T5 encoder you fetch separately remain Apache-2.0.
On the HALLO3 test split Kyutai reports FID 56.9, FVD 195.5 and Sync-C 4.30. In a user study of 400 pairwise comparisons from 27 raters, the model beat Hallo3 on visual quality, lip-sync, reference fidelity and naturalness, and matched or exceeded its own teacher. Yet Kyutai concedes that lip-sync trails the strongest multi-step systems, with Hallo3 at 5.06 and Wan-S2V at 5.42 on Sync-C.
Render the same 30-second English clip with both AURA sizes and with any multi-step baseline you already use. Look at mouth shapes on plosives, at teeth and at identity drift after a minute. Then time the run on your hardware, not ours or Kyutai's. Keep to four sampling steps; one-step sampling degrades sharply. Other languages and non-speech audio are untested.
The same consent and disclosure rules apply as for any face-animation tool. Kyutai's other models are on its provider page, and comparable systems are listed under audio-driven video, including Wan Animate 2.
Compare the model on the work you need to complete.
AURA-5B 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 Snap Inc. Non-Commercial License. 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-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.
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.