Qwen open-sources Qwen3-TTS model family
Alibaba's Qwen team has released open weights for its Qwen3-TTS speech models, which handle voice cloning, voice design and instruction-controlled generation, in 1.7B and 0.6B sizes for 10 languages.
Qwen has announced that the full Qwen3-TTS multi-codebook model series is now open source on GitHub and also available through the Qwen API. As we reported earlier, the family had already gained voice cloning and voice design. The new step is that developers can now run and fine-tune the models themselves.
The release includes five models. At 1.7B parameters there are three: VoiceDesign, which builds a voice from a text description; CustomVoice, which applies instruction-based style control to nine preset timbres; and Base, which clones a voice from three seconds of audio and can be fine-tuned. The 0.6B tier has its own CustomVoice and Base models, but without instruction control. All five support streaming and cover Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.
The models are built on Qwen's Qwen3-TTS-Tokenizer-12Hz speech encoder and a discrete multi-codebook language model, rather than the more common LM+DiT pipeline. Qwen says its "Dual-Track" streaming design can emit the first audio packet after a single input character, with end-to-end latency as low as 97ms.
All benchmark figures come from Qwen. It reports that VoiceDesign beat MiniMax's closed voice-design model on InstructTTS-Eval. In multilingual cloning across 10 languages, it reports an average word error rate of 1.835% and speaker similarity of 0.789, which it says beats MiniMax and ElevenLabs. On continuous 10-minute generation, it reports WER of 2.36% in Chinese and 2.81% in English. Qwen has not said under what license the weights are released.