Qwen3-TTS

weights

open(safetensors on the Hub, ungated)

data

closed(no training corpus named)

code

partial(inference package and fine-tuning scripts)

license

Apache-2.0(OSI)

Qwen3-TTS's weights and code are Apache-2.0, OSI-approved. Alibaba publishes fine-tuning scripts but neither its training data nor the recipe behind the released checkpoints.

• https://huggingface.co/api/models/Qwen/Qwen3-TTS-12Hz-1.7B-Base recorded 2026-09-27

  "license:apache-2.0" tag, "gated":false, with model.safetensors in the file list.

• https://raw.githubusercontent.com/QwenLM/Qwen3-TTS/HEAD/LICENSE recorded 2026-09-27

  LICENSE body: "Apache License" "Version 2.0, January 2004".

• https://raw.githubusercontent.com/QwenLM/Qwen3-TTS/HEAD/README.md recorded 2026-09-27

  "Please refer to [Qwen3-TTS-Finetuning](finetuning/) for detailed instructions on fine-tuning Qwen3-TTS."; no training dataset is named.

Hugging Face downloads of the two most-used checkpoints. The qwen-tts package is marked as not a separate channel because it loads these repositories.

• https://huggingface.co/api/models?author=Qwen&search=Qwen3-TTS recorded 2026-09-27

  Trailing-30-day downloads: Qwen3-TTS-12Hz-1.7B-Base 3,923,955 and Qwen3-TTS-12Hz-1.7B-CustomVoice 2,435,822.

Broad features, voice design and 3-second cloning in 10 languages among them, but the arena entry that most likely corresponds to it sits near the bottom of blind listening tests.

• https://artificialanalysis.ai/text-to-speech/leaderboard recorded 2026-09-27

  Entry "name":"Qwen3 TTS" (creator Alibaba) with "elo":929.65, "releaseDate":"2026-01-22" and "openWeights":false.

• https://raw.githubusercontent.com/QwenLM/Qwen3-TTS/HEAD/README.md recorded 2026-09-27

  The README's news list dates the open release: "2026.1.22".