Qwen3-TTS
Description
Qwen3-TTS is a family of open text-to-speech models from the Qwen team. Per the model cards, it covers 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian) plus dialect voices. There are three kinds of models in 1.7B size (and two in 0.6B): Base clones a voice from about 3 seconds of audio and can be fine-tuned, CustomVoice offers ready-made voices with tone and emotion set by instructions, and VoiceDesign creates a voice from a text description. Streaming output is supported; the authors state a first-packet delay as low as 97 ms. The cards do not state a VRAM requirement and only recommend FlashAttention 2 to save memory. The weights are public on Hugging Face without any acceptance step, released on 2026-01-21 under the Apache-2.0 license, which allows commercial use. The authors offer an official demo Space on Hugging Face, and code and examples are on GitHub. The link on this page leads to the voice-cloning model. Voice cloning should only be used with the speaker's consent. Date of check: 2026-10-19.
Pros and cons
Pros
- 10 languages and dialect voices, per the model cards
- Voice cloning from about 3 seconds of audio
- Voice from a text description and tone control by instructions
- Apache-2.0 license, weights open without an acceptance step
- Smaller 0.6B models and an official Space
Cons
- No VRAM requirement in the cards
- The family is split into several separate models by task
- Latency and quality statements are the authors' own
- Not on the public speech arena as open weights, so quality is our expert estimate
Key facts
- License
- Apache-2.0
- Commercial use
- Allowed
- Developer
- Qwen (Alibaba)
- Interface languages
- English
- Checked on
- Oct 19, 2026