Can AuK be used in a commercial product?

AuK is released under the MIT license, one of the most permissive open-source licenses, so commercial use, modification and redistribution are all allowed with attribution. That is a notable difference from many open speech models, which ship under research-only or non-commercial terms and cannot be put into a paid product without a separate agreement.

What is the difference between AuK and AuK-Flash?

AuK-Flash is a distilled version of the base AuK model. The technical report says it reaches near-teacher generation quality in four sampling steps, without classifier-free guidance at inference, for a 4.5x wall-clock speedup. Base AuK stays the higher-quality option; AuK-Flash is aimed at latency-sensitive or high-volume work.

What kinds of speech edits can AuK make?

AuK handles content edits such as rewriting spoken words and lyrics, acoustic edits including pitch, speed and volume, and paralinguistic edits covering emotion, timbre, accent and nonverbal sounds. It also does enhancement work: denoising, dereverberation and separating overlapping speakers. Each is requested as a natural-language instruction rather than a separate model.

Where do I download the AuK weights?

AuK weights are published on Hugging Face as tencent/AuK and tencent/AuK-Flash, and on ModelScope as Tencent-Hunyuan/AuK and Tencent-Hunyuan/AuK-Flash. The GitHub repository holds the inference code, which installs with pip into a Python 3.10 environment and exposes an auk-infer command-line entry point.

Tencent Hunyuan · 2026-09-08 · major

AuK — Tencent's open speech model generates and edits audio by instruction

AuK is an MIT-licensed 1.5B speech model from Tencent Hunyuan, Shanghai Jiao Tong University and the Shanghai Innovation Institute. It generates and edits speech from plain-language instructions. Code and weights shipped on September 8, 2026.

Hugging Face model card banner for Tencent's AuK speech model

One 1.5B open model that writes, rewrites, cleans and restyles speech from a sentence of instructions.

Key specs

Flash inference steps	4
Flash speedup	4.5x

Quick facts

Maker	Tencent Hunyuan, Shanghai Jiao Tong University, Shanghai Innovation Institute
Size	1.5B parameters
License	MIT
Variants	AuK and AuK-Flash
Training data	3.03B instruction-audio instances, 1.95M hours
Weights	Hugging Face and ModelScope

What is it?

Speech generation and speech editing usually need separate models; AuK puts both behind one interface where you pass an audio clip and a plain-language instruction. The 1.5B model comes from Tencent Hunyuan with Shanghai Jiao Tong University and the Shanghai Innovation Institute, and covers five task families: generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing. Code and weights are MIT-licensed.

How does it work?

AuK combines a diffusion transformer with what the paper calls layer-fusion weights, and uses Qwen2.5-Omni-3B as the multimodal encoder that reads the instruction alongside the audio. Training covered roughly 3.03 billion instruction-audio pairs drawn from about 1.95 million hours of supervision. A distilled sibling, AuK-Flash, reaches near-teacher quality in four sampling steps with no classifier-free guidance at inference.

Why does it matter?

Anyone building voice products has been chaining a TTS model, a denoiser, a separator and a prosody tool, each with its own license. AuK collapses that chain into one MIT-licensed checkpoint, so changing a speaker's emotion, fixing a word, removing room echo or splitting two voices all become the same call. The MIT terms also make it usable in commercial products without a negotiation.

Who is it for?

speech and audio engineers

Frequently asked questions

Can AuK be used in a commercial product?: AuK is released under the MIT license, one of the most permissive open-source licenses, so commercial use, modification and redistribution are all allowed with attribution. That is a notable difference from many open speech models, which ship under research-only or non-commercial terms and cannot be put into a paid product without a separate agreement.
What is the difference between AuK and AuK-Flash?: AuK-Flash is a distilled version of the base AuK model. The technical report says it reaches near-teacher generation quality in four sampling steps, without classifier-free guidance at inference, for a 4.5x wall-clock speedup. Base AuK stays the higher-quality option; AuK-Flash is aimed at latency-sensitive or high-volume work.
What kinds of speech edits can AuK make?: AuK handles content edits such as rewriting spoken words and lyrics, acoustic edits including pitch, speed and volume, and paralinguistic edits covering emotion, timbre, accent and nonverbal sounds. It also does enhancement work: denoising, dereverberation and separating overlapping speakers. Each is requested as a natural-language instruction rather than a separate model.
Where do I download the AuK weights?: AuK weights are published on Hugging Face as tencent/AuK and tencent/AuK-Flash, and on ModelScope as Tencent-Hunyuan/AuK and Tencent-Hunyuan/AuK-Flash. The GitHub repository holds the inference code, which installs with pip into a Python 3.10 environment and exposes an auk-infer command-line entry point.

Try it

auk-infer --audio input.wav --instruction "Your instruction" --output out.wav