Qwen releases Qwen3-TTS, Apache 2.0 voice cloning in 10 languages
Qwen3-TTS is a family of open TTS models, 0.6B and 1.7B, with 3-second voice cloning, voice design and streaming, under Apache 2.0.
Alibaba's Qwen team released Qwen3-TTS, a family of open text-to-speech checkpoints. The card for the 1.7B Base model lists Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The Base model clones a voice from "a 3-second rapid voice clone" of reference audio, plus an optional transcript. Sibling checkpoints cover a 1.7B VoiceDesign model, 1.7B and 0.6B CustomVoice models with nine preset timbres, a 0.6B Base model and a tokenizer.
The card cites end-to-end synthesis latency "as low as 97ms" for the series' dual-track streaming design. It does not say that figure applies to this particular checkpoint. On the Seed-TTS test set the card reports a word error rate of 0.77 for Chinese and 1.24 for English, and it lists CosyVoice 3 ahead on Chinese at 0.71.
The licence is Apache 2.0. The card does not give a release date; the date here, 22 January 2026, is the date of the linked technical report on the card. The Hugging Face repo was created a day earlier.
It does what the card says. The latency figure comes with a caveat the headline omits.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Weights for several checkpoints are on Hugging Face under Apache 2.0, with vLLM-Omni support noted.
- Useful 4: Cloning, preset voices and voice design cover most TTS tasks in ten languages.
- Notable 4.5: Several checkpoints, a tokenizer and vLLM-Omni support make it a ready-to-use family.
- Novel 3: Dual-track streaming and a family layout are tidy, but the idea is incremental.
Comments