Tencent Open-Sources AuK, a 1.5B-Parameter Speech Generation and Editing Model

Tencent has released AuK, a 1.5B-parameter open-source foundation model for speech generation and editing, with code and weights publicly available on Hugging Face and ModelScope as of September 9, 2026.

What AuK Does

AuK is trained on millions of hours of diverse audio data, according to Tencent, and unifies a wide range of speech tasks under a single natural-language instruction interface. Supported capabilities span four categories:

• Speech generation: zero-shot TTS (voice cloning from reference audio) and instruct TTS (voice generation from text description alone)
• Content editing: rewriting spoken content (replace, insert, remove text) and lyric editing that preserves melody and voice
• Acoustic and paralinguistic editing: pitch, speed, and volume adjustment; emotion and timbre transfer; accent removal; nonverbal sound editing (breaths, laughs, coughs); whisper conversion
• Enhancement and separation: speech enhancement (denoising, dereverberation), speech separation, music/vocal separation, and target speaker extraction

The model ships in two variants: AuK, the base model optimized for generation quality, and AuK-Flash, a distilled version built for fast 4-step inference. Both rely on a Qwen2.5-Omni-3B MLLM encoder loaded separately at runtime alongside a VAE component.

Access and Licensing

AuK is released under the MIT license, with weights downloadable via Hugging Face CLI or ModelScope. The SGLang Omni team has already implemented what Tencent calls "Day 0 support" for serving the model. As of publication, no inference provider has deployed AuK on a hosted API — the model must be self-hosted, so no per-token or per-character pricing applies.

Tencent has not published specific benchmark scores in the model card comparing AuK against other open or closed speech models; the accompanying technical report (arXiv:2609.08936) presumably contains detailed evaluation data but was not included in the released materials reviewed here.

What This Means

AuK adds to a growing field of open-weight audio foundation models but distinguishes itself by folding generation, editing, enhancement, and separation into one instruction-driven interface rather than shipping separate specialized models for each task. That design choice — plus the MIT license and a distilled fast-inference variant — makes AuK immediately usable for developers building voice applications who need to self-host rather than rely on a commercial TTS API.

The absence of head-to-head benchmark numbers against competitors like ElevenLabs, OpenAI's audio models, or other open releases makes it hard to independently verify quality claims at this stage. Engineers evaluating AuK for production use should treat the 1.5B parameter count and multi-task claims as a starting point, and validate output quality against their own use cases before deployment. The requirement to separately download a 3B-parameter Qwen encoder alongside the AuK weights also means total resource requirements are higher than the AuK checkpoint size alone suggests.