Tencent Open-Sources AuK for 16 Audio Tasks

Tencent’s Hunyuan team, in collaboration with Shanghai Jiao Tong University and the Shanghai Innovation Institute, has released AuK, an open-source audio foundation model. This 1.5-billion-parameter model is designed to handle 16 distinct audio tasks through a unified natural-language interface. AuK relies on a separately downloaded Qwen2.5-Omni-3B model for semantic conditioning. Its architecture features a 50 Hz audio VAE and a hybrid MMDiT plus single-stream DiT rectified-flow transformer.

To achieve its versatile capabilities, the development team trained AuK on 3.03 billion instruction-audio pairs and 1.95 million hours of audio across five task families, followed by reinforcement learning from human feedback post-training. For developers requiring faster performance, Tencent also released AuK-Flash, a distilled variant that delivers a 4.5x speedup and enables four-step inference when classifier-free guidance is disabled.

For AI practitioners, AuK replaces traditional, fragmented speech pipelines that require stitching together separate services for text-to-speech, voice conversion, denoising, and editing. Instead, developers can use a single chat-style API to perform zero-shot text-to-speech, content editing, source separation, and acoustic adjustments. This unified approach significantly reduces the integration and routing code needed to connect specialist systems. Users simply provide an instruction alongside optional source or reference clips to generate the desired waveform.

The model is released under the permissive MIT license, making it highly accessible for commercial and research applications. The weights are currently available on Hugging Face and ModelScope. Additionally, the project includes support for Gradio, ComfyUI, and custom fine-tuning, allowing developers to easily integrate the model into existing creative and conversational AI workflows.