AuK is a 1.5B-parameter foundation model for speech generation and editing that Tencent's Hunyuan team open-sourced on September 9, 2026, alongside a distilled variant called AuK-Flash. Rather than shipping separate models per audio task, AuK exposes zero-shot TTS, instruction-based TTS, content editing, acoustic editing, paralinguistic editing, speech enhancement, and source separation through one natural-language instruction interface: an instruction plus optional source or reference audio, handled by the same auk-infer CLI call. Zero-shot TTS clones a reference voice to speak new target text, while instruct TTS synthesizes a voice purely from a text description with no reference audio. Content editing rewrites what is said in a recording -- replacing, inserting, or removing words, or rewriting sung lyrics while preserving melody and voice. Acoustic editing adjusts pitch by semitones, speaking rate, or volume by decibels. Paralinguistic editing changes emotion, timbre, or accent, adds or removes nonverbal sounds like breaths, laughs, and coughs, and converts speech to and from a whisper register. The enhancement and separation tasks denoise recordings, extract a target speaker identified by what they say, or split a singing voice out of a music mix. AuK's architecture pairs a diffusion transformer with a Qwen2.5-Omni-3B encoder that reads the instruction and conditions generation on it, downloaded alongside the checkpoint. AuK-Flash is a 4-step distilled version for faster inference at fixed CFG=0, trading quality for latency, while the base model exposes configurable NFE and guidance. Both variants run from the command line, a Gradio web demo, or as ComfyUI custom nodes, and download from Hugging Face and ModelScope. The Gradio demo adds an optional LLM-driven Prompt Enhancer that turns a free-form request into a ready-to-run auk-infer command, calling Tencent Cloud's recording-file recognition for audio-backed prompts or falling back to a local SenseVoiceSmall model when no cloud ASR credentials are set. The project ships fine-tuning support and a cookbook of instruction templates with CLI and Python examples for every task, and is released under the MIT license.