01
Choose Base or AuK-Flash
Use AuK Base for the configurable quality-focused path and AuK-Flash when you want Tencent's fixed four-step distilled checkpoint. Actual latency still depends on hardware, duration and the rest of the runtime.
Do not convert the four-step design into an unsupported universal real-time claim. Benchmark both variants on the exact speech task you plan to ship.
02
Install the official environment
Tencent documents cloning Tencent-Hunyuan/AuK, creating a Python 3.10 environment with uv or Conda and installing the repository. Optional extras add Gradio, ComfyUI or fine-tuning dependencies.
Install a platform-appropriate PyTorch stack when needed, then use the live repository instructions rather than pinning a launch-day dependency version in perpetuity.
03
Download the required weights
The official Hugging Face path downloads tencent/AuK for Base, tencent/AuK-Flash for Flash and Qwen/Qwen2.5-Omni-3B as the encoder. The expected checkpoint layout keeps these as separate directories.
Text-encoder keys can appear absent from the diffusion checkpoint because the encoder is loaded separately. Do not treat that documented separation as a corrupt checkpoint.
04
Run and plan hardware cautiously
The CLI always uses a natural-language instruction and can add source or reference audio depending on the task. Gradio is available through the documented optional extra, while Prompt Enhancer requires an OpenAI-compatible LLM configuration.
Tencent does not publish one universal minimum-VRAM number, so measure memory with your dtype, loaded variant and audio duration. For task coverage and licensing, return to the main AuK guide; for node workflows see the ComfyUI guide.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.