AuK

Tencent's 1.5B instruction-driven model for speech generation, editing, enhancement and separation, plus a 4-step Flash. MIT license.

In short. AuK is a speech model that follows written instructions. It can read text in a cloned or described voice, and it can also change a recording you already have, for example replacing a few words, making the voice calmer, removing an accent, cleaning up noise or pulling one speaker out of a mix. It is free to use, including commercially, under the MIT license.

Summary

Tencent's Hunyuan team open-sourced AuK on 9 September 2026 under the MIT license, with a technical report, arXiv 2609.08936. It is a 1.5-billion-parameter diffusion transformer trained on millions of hours of audio, and it treats speech generation and editing as one problem, since every task is a written instruction plus optional source audio. Besides zero-shot and instruction-based text-to-speech, where you describe a voice with no reference clip, it can rewrite what is said in a recording by replacing, inserting or removing words, edit the lyrics of a sung recording, change pitch, speed and volume, change emotion, timbre or accent, add or remove breaths and laughs, convert between whisper and normal speech, denoise and dereverberate speech, and separate speakers, singing voices or a target speaker.

There are two checkpoints, the base AuK and AuK-Flash, a distilled version that runs 4 fixed steps without guidance. Each is a 6.1 GB FP32 file plus a 0.64 GB VAE, and the pipeline also loads Qwen2.5-Omni-3B as its multimodal encoder, which you download separately. The README measures peak memory on an A800 in bf16 at about 25 GiB, which falls to about 17 GiB with CPU offload. It offers a command-line tool, a Gradio demo, ComfyUI nodes with a 30-second limit on source plus target audio, a Python API and fine-tuning code, and vLLM-Omni and SGLang-Omni can serve it. The README also lists MLX support for Apple silicon on a branch and a community GGUF port in audio.cpp. An optional prompt enhancer turns a free-form request into the exact instruction and duration, and it needs an OpenAI-compatible language model.

The card shows its benchmark results only as a chart, so we quote none of them, and it gives no list of supported languages, although the demos include English and Chinese. The ICASSP 2027 Audio Editing Challenge uses AuK as its single-model baseline. In the 30 days to 11 November 2026, Hugging Face counted about 1,300 downloads of the base model and about 700 of AuK-Flash.

Good for

  • Editing speech in an existing recording from a written instruction, such as changing words, pitch, speed, volume, emotion or accent
  • Cleaning up audio, with denoising, dereverberation, speaker separation and singing-voice extraction
  • Zero-shot voice cloning, or generating speech from a description of a voice
  • Research and commercial work that needs an MIT-licensed speech model

Watch out for

  • Qwen2.5-Omni-3B has to be downloaded separately, and Tencent measured about 25 GiB of peak GPU memory, or about 17 GiB with CPU offload
  • The benchmarks appear only as an image, and the quality claims are Tencent's
  • Free-form requests work best through the prompt enhancer, which needs an external OpenAI-compatible language model
  • Cloning or editing someone's voice can be used to fake their speech, and the MIT license does not set consent rules
  • No supported-language list is published, and the demos cover English and Chinese

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of AuK. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
AuK (base) FP32 1.5B 6.8 GB 18 GB Gaming GPU
AuK-Flash (4-step distilled) FP32 1.5B 6.8 GB 18 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 18 GB of GPU memory.

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

See also

Frequently asked questions

What is AuK?
AuK is a speech model that follows written instructions. It can read text in a cloned or described voice, and it can also change a recording you already have, for example replacing a few words, making the voice calmer, removing an accent, cleaning up noise or pulling one speaker out of a mix. It is free to use, including commercially, under the MIT license.
Can I use AuK commercially?
Yes. AuK is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is AuK open source?
AuK is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run AuK on my own computer?
Yes, on a good graphics card. The smallest variant, AuK (base), needs about 18 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was AuK released?
AuK was released by Tencent on September 9, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
VoxCPM2 OpenBMB 2.29B – Apache 2.0 (Permissive) 8 GB Laptop or phone
OmniVoice Xiaomi 613M – CC BY-NC (Non-commercial) 3.75 GB Laptop or phone
Qwen3-TTS Alibaba Qwen 910M – 1.92B – Apache 2.0 (Permissive) 3.68 GB Laptop or phone
Fun-CosyVoice 3 Tongyi Lab Speech Team 500M – Apache 2.0 (Permissive) 6.21 GB Laptop or phone