Qwen3-TTS

Alibaba's Apache 2.0 text-to-speech models in 1.7B and 0.6B sizes: 10 languages, 3-second voice cloning, streaming at about 97 ms.

In short. Qwen3-TTS reads text aloud in a natural voice. You can pick one of nine built-in voices, describe the voice you want in words, or clone a voice from three seconds of audio. It speaks ten languages, can be told how to sound (calm, excited, fast, slow), and starts speaking almost immediately, so it suits live assistants. The license is Apache 2.0.

Summary

Alibaba's Qwen team released Qwen3-TTS on 22 January 2026 under the Apache 2.0 license. It is a family of text-to-speech models in two sizes, 1.7B and 0.6B, with five open checkpoints. VoiceDesign (1.7B) creates a new voice from a written description. CustomVoice (1.7B and 0.6B) offers nine premium built-in voices covering combinations of gender, age, language and dialect, with style control through instructions. Base (1.7B and 0.6B) can clone a voice from about three seconds of reference audio and is the checkpoint to fine-tune.

The models cover ten languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian) plus several dialect voice profiles. Qwen says they adjust tone, speaking rate and emotion from natural-language instructions and the meaning of the text, and cope better with noisy input text. Speech is represented by a separate 12 Hz speech tokenizer (Qwen3-TTS-Tokenizer-12Hz, a 0.7 GB download), and the models are discrete multi-codebook language models rather than a language model followed by a diffusion stage. A dual-track design supports both streaming and non-streaming output in one model: the first audio packet can be produced after a single character, with end-to-end latency as low as 97 ms, according to Qwen.

Each 1.7B checkpoint is about 4.5 GB and each 0.6B checkpoint about 2.5 GB, plus the tokenizer. The qwen-tts Python package loads them, and vLLM is supported; Qwen recommends FlashAttention 2 to reduce GPU memory. The latency and quality comparisons in Qwen's technical report are self-reported.

Good for

  • Voice assistants and agents that need low-latency streaming speech
  • Multilingual narration, audiobooks and video voice-overs in 10 languages
  • Creating a custom voice from a text description, without recordings
  • Cloning a voice you have the right to use from a short sample, and fine-tuning the Base model

Watch out for

  • Cloning a real person's voice without their permission can be illegal or harmful, and Apache 2.0 does not give you that right
  • Only ten languages are supported; check quality for your language and dialect before relying on it
  • The 97 ms latency is Qwen's best case and depends on your hardware and serving stack
  • Quality and latency comparisons come from Qwen's own report

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Qwen3-TTS. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
0.6B Base (voice cloning, fine-tuning) BF16 910M 3.2 GB 3.68 GB Laptop or phone
0.6B CustomVoice (built-in voices) BF16 910M 3.2 GB 3.68 GB Laptop or phone
1.7B Base (voice cloning, fine-tuning) BF16 1.92B 5.2 GB 5.98 GB Laptop or phone
1.7B CustomVoice (built-in voices) BF16 1.92B 5.2 GB 5.98 GB Laptop or phone
1.7B VoiceDesign (voice from a description) BF16 1.92B 5.2 GB 5.98 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Qwen3-TTS?
Qwen3-TTS reads text aloud in a natural voice. You can pick one of nine built-in voices, describe the voice you want in words, or clone a voice from three seconds of audio. It speaks ten languages, can be told how to sound (calm, excited, fast, slow), and starts speaking almost immediately, so it suits live assistants. The license is Apache 2.0.
Can I use Qwen3-TTS commercially?
Yes. Qwen3-TTS is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Qwen3-TTS open source?
Qwen3-TTS is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Qwen3-TTS on my own computer?
Yes. The smallest variant, 0.6B Base (voice cloning, fine-tuning), needs about 3.68 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Qwen3-TTS released?
Qwen3-TTS was released by Alibaba Qwen on January 22, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 8, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
MOSS-TTS OpenMOSS 4.55B – 8.49B – Apache 2.0 (Permissive) 10 GB Laptop or phone
LFM3-Audio-2B Liquid AI 2B 32K LFM Open 1.0 (Conditional) 4.6 GB Laptop or phone