How to Run tencent/AuK Locally: Zero-Shot TTS, Voice Cloning and Speech Editing

tencent/AuK is a 1.5 billion parameter open source text-to-speech model, released August 18, 2026 under the MIT license. One natural-language instruction interface handles zero-shot voice cloning, speech editing, pitch and speed changes, and source separation, but it also loads a separate 3 billion parameter Qwen2.5-Omni encoder at runtime, so the real footprint running the model is closer to 4.5 billion parameters.

Written by Priya Raghunathan, local-AI hardware reviewer. I check parameter counts and hardware requirements against the primary source before recommending an install path, whether the model ships an official VRAM figure or not.

A three-week-old model, already in the trending top 20

AuK entered Hugging Face's trending top 20 for the text-to-speech pipeline within about three weeks of its August 18, 2026 release, an unusually fast climb for a model from a research lab rather than a consumer app. At time of writing it has 20 likes and 30 downloads on the Hugging Face repo itself, both still small numbers, plus 166 stars on the Tencent-Hunyuan/AuK GitHub repo and 11 likes on the tencent/AuK demo Space. Read that combination as early attention from people who build TTS pipelines, not yet a mainstream download count.

Specs at a glance

Hardware: no published VRAM figure, so here's the math

Neither the model card nor the GitHub repo lists a VRAM requirement or an inference speed benchmark. What can be calculated: AuK's own 1.5 billion parameters plus the 3 billion parameter Qwen2.5-Omni-3B encoder it depends on comes to about 4.5 billion parameters resident in memory at once. At 2 bytes per parameter in bf16, that's roughly 9 GB of VRAM or unified memory for weights alone, before any KV cache or activation overhead from the diffusion sampling steps.

Supported tasks: one instruction interface, five categories

The thing that sets AuK apart from a typical TTS model is that generation, editing, cleanup and separation all run through the same natural-language instruction, rather than five separate tools.

Install it

AuK is a source install, not a pip package. The GitHub repo uses uv with Python 3.10 as its primary path, with conda as an alternative.

Download the weights

Because AuK relies on the Qwen2.5-Omni-3B encoder at inference time, plan on two downloads, not one.

Run zero-shot TTS

Point the CLI at a reference audio clip and an instruction, and AuK generates speech in that voice.

auk-infer \
    --audio assets/demo-input-audio/zero-shot-tts/ref.wav \
    --instruction "Say the following with the same voice: 'Ladies and gentlemen, it's an honor to have the opportunity to address such a distinguished audience'" \
    --output out_tts.wav \
    --gen_seconds 6.0

The Python API takes the same reference audio and instruction as a chat-style message list, which is closer to what you'd wire into an app.

from auk.infer.infer_auk import AukInfer, save_audio

engine = AukInfer("ckpts/AuK/config.yaml", "ckpts/AuK/auk_base.safetensors")
messages = [{
    "role": "user",
    "content": [
        {"type": "text", "text": "Say: 'Hello world' with a clear voice"},
        {"type": "audio", "audio": "ref.wav"},
    ],
}]
audio, sr = engine.generate(messages, gen_seconds=3.0)
save_audio(audio, sr, "output.wav")

AuK is now open-source. Code and model weights are publicly available.

AuK vs AuK-Flash

AuK, the model in this repository, is the base model built for quality. AuK-Flash is a distilled version of the same model built for speed, generating audio in four inference steps instead of the base model's full diffusion sampling schedule. Start with the base model if you're evaluating output quality; switch to AuK-Flash once you know the workflow and want faster turnaround.

License and who it's for

Both AuK and AuK-Flash are MIT licensed, with no revenue cap and no required attribution. That's as permissive as it gets. What MIT doesn't buy you is maturity: at 30 downloads and three weeks old, the tooling, community troubleshooting and third-party integrations are thin compared to an established local TTS model. If you need something that already has a well-worn install path today, that's a reason to start with Kokoro-82M or Breeze TTS 2 instead and come back to AuK once its ecosystem catches up to its feature list.

FAQ

What is tencent/AuK used for?

Zero-shot voice cloning and instruction-guided text-to-speech, plus speech content editing, pitch and speed changes, emotion and timbre changes, and audio cleanup tasks like denoising, speech separation and speaker extraction, all through one natural-language instruction interface instead of separate models per task.

How many parameters does AuK have?

The model card lists AuK's base model at 1.5 billion parameters. It also loads a separate 3 billion parameter Qwen2.5-Omni-3B encoder at inference time to read text and reference audio, so the practical memory footprint is closer to 4.5 billion parameters combined.

Does AuK need a GPU?

Tencent hasn't published an official VRAM figure. Based on the combined 4.5 billion parameter footprint at bf16 precision, expect to need roughly 9 GB of VRAM or unified memory for weights alone, which puts it comfortably on a 16 GB or 24 GB consumer GPU rather than an 8 GB card.

What's the difference between AuK and AuK-Flash?

AuK is the base model, built for quality. AuK-Flash is a distilled version of the same model, built for speed, generating audio in four inference steps instead of the base model's full diffusion sampling schedule.

Is tencent/AuK free to use commercially?

Yes. Both the AuK and AuK-Flash weights are released under the MIT license, which has no revenue cap and no required attribution.

Where to go from here

• Want a much smaller, more battle-tested TTS model to start with? See how to run Kokoro-82M locally.
• Comparing voice clone and voice design workflows? How to run Breeze TTS 2 locally covers a similar feature set with a smaller, more established model.