
AURA-5B is an audio-driven talking-head video model from Kyutai, published in November 2026. It takes one portrait image plus a speech audio stream and generates 720p video chunk by chunk with no audio look-ahead. It is a 4-step causal diffusion student distilled from the 50-step TalkVerse-S2V-5B teacher, with 5.7B parameters, and runs at 9.7 fps end-to-end on a single H100. Weights are released under the Snap Inc. Non-Commercial License for research use only.
A workable 5.7B-parameter dense video generator from kyutai. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | SS | 4.0 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 4.0 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 4.0 GB |
| AMD Instinct MI300XAMD | SS | 4.0 GB |
| AMD Instinct MI325XAMD | SS | 4.0 GB |
Cheapest current cloud rentals with at least 4 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.05 |
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.07 |
NVIDIA Tesla V100 16GBVast.ai · On-Demand · 16 GB VRAM | $0.07 |
NVIDIA GeForce RTX 4060 TiVast.ai · Spot · 8 GB VRAM | $0.08 |
NVIDIA GeForce RTX 3080Vast.ai · Spot · 10 GB VRAM | $0.08 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
AURA-5B is an audio-driven streaming avatar model developed by the French open-science lab Kyutai. Designed to turn a single reference portrait and a live speech audio stream into 720p talking-head video, it solves one of the primary bottlenecks in interactive video generation: latency. Traditional video diffusion pipelines require multi-second audio buffers and 30 to 50 denoising steps before yielding a single frame. AURA-5B reduces this to 4 causal denoising steps per chunk with zero audio look-ahead, generating video sequentially as the speech arrives.
Built on top of Snap's TalkVerse-S2V-5B (which itself builds on Wan2.2-TI2V-5B), the model operates as a dense 5.7B parameter diffusion transformer (DiT). By using causal distillation, Kyutai achieved a 17x end-to-end acceleration over the original 50-step teacher model, clocking 9.7 frames per second (fps) at 720p resolution on a single NVIDIA H100 GPU. Released under the Snap Inc. Non-Commercial License for research applications, AURA-5B represents a critical transition point for practitioners looking to bridge speech agents and photorealistic visual synthesis.
While smaller models like AURA-1.3B prioritize real-time throughput at lower resolutions, AURA-5B is optimized for visual fidelity and facial detail. It replaces legacy 2D warping pipelines with a fully generative diffusion architecture capable of natural eye blinks, accurate lip synchronization, and micro-expressions driven directly by speech features.
The architecture of AURA-5B centers on a dense 5.7B parameter diffusion transformer trained through progressive distillation. Unlike mixture-of-experts (MoE) designs where only a subset of parameters activates per token, a dense architecture requires all 5.7 billion parameters to participate in every calculation across every denoising step. This demands sustained memory bandwidth and raw compute throughput from the host GPU.
The pipeline consists of three core components:
wav2vec2-large-xlsr-53-english. Because the model is strictly causal, the audio encoder processes incoming audio stream chunks without relying on future frames or bidirectional audio context.Checkpoints are distributed in two configurations: a standalone merged DiT checkpoint (*_merged) stored as fp32 safetensors weighing 22.8 GB, and a training-oriented LoRA checkpoint (rank 128, alpha 128) containing the delta weights alongside the trained audio projection and output modules.
AURA-5B is engineered for synchronous, speech-driven character animation. It takes a static portrait, reads raw waveform audio, and synthesizes 25 fps playback video aligned with the audio track.
Practical deployments focus on scenarios requiring high visual clarity where batched generation is impractical:
Because the underlying audio encoder is wav2vec2-large-xlsr-53-english, the model achieves its highest lip-sync fidelity on English speech. Phoneme-to-viseme alignment degrades on non-English inputs unless the audio encoder is replaced or adapted. Furthermore, commercial deployments are restricted by the Snap Inc. Non-Commercial License, confining local deployments strictly to academic and non-commercial research environments.
Deploying AURA-5B on local workstations requires planning for the combined memory footprint of the 5.7B DiT, the wav2vec2 feature extractor, the spatial-temporal VAE, and the rolling key-value cache.
Unlike text-only language models that run cleanly inside Ollama or llama.cpp, AURA-5B is an image-and-audio-to-video diffusion pipeline that runs via custom PyTorch pipelines. Practitioners who normally search for AURA-5B tokens per second must instead measure throughput in video frames per second (fps) generated per second of real time.
The best quantization for AURA-5B in consumer environments is FP8 (specifically FP8-E4M3 or FP8-E5M2 formats supported by PyTorch and modern CUDA architectures). This halves the DiT parameter footprint from 11.5 GB down to approximately 5.8 GB, allowing developers exploring how to run 5.7B model on consumer GPU setups to easily fit the full pipeline onto an RTX 4090 or an Apple Silicon M4 Max (using unified memory).
Regarding AURA-5B performance:
torch.compile). The DiT itself runs at 19.7 fps, while the VAE decode step executes at 22.7 fps.torch.compile and FP8 scales throughput closer to 5.5 to 6.5 fps, which remains below the 25 fps target needed for native real-time playback.The best GPU for AURA-5B if you need real-time or near-real-time generation is an enterprise-class accelerator like the NVIDIA H100, H200, or a dual L40S configuration. To run AURA-5B locally, clone the kyutai-labs/aura repository, install dependencies via uv, and launch the inference script with --compile enabled for CUDA graph acceleration.
Choosing AURA-5B involves balancing visual resolution against hardware throughput when compared to alternative local models:
For researchers evaluating a local AI model 5.7B parameters 2026 release, AURA-5B delivers state-of-the-art causal video generation, trading off the low hardware requirements of 2D warp engines for authentic generative visual quality.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every kyutai model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.