[MODEL REQUEST] Qwen3-TTS adaptation for Qualcomm devices with NPU

[MaTwickenham]

Is your feature request related to a problem? Please describe.
I am working on adapting Qwen3-TTS for on-device inference on Qualcomm Snapdragon devices, especially with NPU / HTP acceleration.

Qwen3-TTS is a strong open-source TTS model, but deploying it efficiently on mobile devices requires model decomposition, export-path adaptation, backend placement, and runtime optimization. I have already built an initial Qualcomm-oriented prototype and achieved around RTF ≈ 1 in my local test setup. I would like to discuss whether this work can be aligned with qualcomm/ai-hub-models and potentially upstreamed as an official model contribution.

Details of model being requested

• Model name: Qwen3-TTS 0.6B series
• Source repo link: https://github.com/QwenLM/Qwen3-TTS
• Research paper link [If applicable]: https://arxiv.org/abs/2601.15621
• Model use case: On-device text-to-speech, low-latency speech generation, custom voice generation, and privacy-preserving TTS on Qualcomm Snapdragon devices.

Additional context for requested model
I have completed an initial Qualcomm-oriented adaptation of Qwen3-TTS. The current work includes:

• Model component decomposition for deployment-oriented execution
• Model architecture adaptation for Qualcomm runtime compatibility
• Mixed-precision quantization across different model components
• Experiments with different compute backends, including CPU and NPU / HTP paths

Based on this preliminary prototype, I would like to collaborate with the Qualcomm AI Hub Models maintainers to further improve and upstream this work. I am especially looking for guidance on model structure optimization, quantization strategy, backend placement, and the preferred integration path for making this model work well on Qualcomm devices.

[gustavla]

Thanks for reaching out! We are actually in the process of trying to make it easier to make stand-alone repositories with AI Hub Models-style Qualcomm-ready onboarded model recipes. This will open the door to 3P hosted repositories (and the upstream path in to AI Hub Models easy as well).

So, currently we don't have the infrastructure to take contributions like this (code is totally open, but CI infra is currently not), but stay tuned and this may change in the future.