AuK-Flash: 4-Step Distilled Speech Model by Tencent

ComfyUI Wiki

AuK-Flash is the distilled variant of Tencent's AuK speech model, running fixed 4-step inference with CFG=0 for fast generation and editing in ComfyUI.

A

AuK-Flash

TTSDistilledFast InferenceSpeech Editing

Distilled variant of Tencent's AuK speech model. Uses fixed 4-step inference with CFG=0 for faster generation and editing while keeping the full task coverage: TTS, voice cloning, speech editing, enhancement, and separation.

HF RepoGitHubComfyUI Guide
DeveloperTencent Hunyuan + Shanghai Jiao Tong University
Release Date2026-09-09
Architecture1.5B diffusion transformer, distilled (NFE=4)
LicenseMIT
Parameters1.5B
Inference4 fixed steps, CFG=0

Capabilities

AuK-Flash shares the same task coverage and instruction interface as the AuK base model: zero-shot and instruct TTS, content, acoustic, and paralinguistic editing, speech enhancement, and source separation. The trade-off is speed over maximum quality.

ComfyUI Usage

Load auk_flash.safetensors in AuK Model Loader and set:

  • nfe_steps=4
  • cfg_strength=0
  • sway_sampling_coef=-1

The same Qwen2.5-Omni-3B encoder and the 30-second source-plus-target limit apply. See the official ComfyUI guide for details.

Guides and workflows related to this model series.

No articles found.

Comments

Sign in with GitHub to join the discussion.

Loading comments…