Qwen3-TTS

Free plan APILinuxSelf-hostedWeb

What Qwen3-TTS is

Qwen3-TTS is an open-source series of text-to-speech models from Alibaba Cloud’s Qwen team. It generates speech in streaming or non-streaming mode and supports voice design from natural-language descriptions, voice cloning from reference audio, and natural-language voice control. Released models include 0.6B and 1.7B parameter versions for custom voice or base tasks, plus a 1.7B voice design model. The listed language count is 10, and WAV is an export format. The project reports end-to-end synthesis latency as low as 97 ms. Users can install the qwen-tts Python package, load model weights from Hugging Face or ModelScope, or launch the local Gradio interface with qwen-tts-demo. The project recommends a fresh Python 3.12 environment and demonstrates loading models on CUDA; FlashAttention 2 is recommended for reducing GPU memory use on compatible hardware. Alibaba Cloud DashScope provides linked real-time APIs for custom voice, cloning, and voice design; API use requires an API key. The repository uses the Apache-2.0 license and permits commercial use. vLLM-Omni supports offline inference; online serving is planned for later.

Who it suits

Qwen3-TTS suits developers and teams seeking open-source speech generation, voice design, or voice cloning. Its local installation and API options serve different deployment needs.

What it does well

Supports streaming and non-streaming speech generation.
Voice cloning uses reference audio.
Apache-2.0 license permits commercial use.
Local Gradio interface is available.

What to know first

Local model setup demonstrates CUDA loading.
FlashAttention 2 requires compatible hardware.
vLLM-Omni online serving is planned, not currently supported.

Plans

Plan	Billed and limits	Price
Qwen3-TTS open-source models	The project is released under the Apache-2.0 license; no price is stated. · 0.6B and 1.7B models · 10 languages	Free

As github.com lists them, read 7 Oct 2026. The maker's own page has the final word.

Getting started

Create a fresh Python 3.12 environment.
Install the `qwen-tts` Python package from PyPI.
Download model weights from Hugging Face or ModelScope, or load them by model ID.
For local use, load a model in the demonstrated CUDA setup; consider FlashAttention 2 on compatible hardware.
Launch the local Gradio interface with `qwen-tts-demo`.
For DashScope API use, configure an API key and install the latest DashScope SDK when using that SDK.

Verdict

Qwen3-TTS offers local and API paths for speech generation, voice design, and cloning, with commercial use permitted under Apache-2.0. Check the documented hardware and API-key requirements for the route you plan to use.

Questions

Is Qwen3-TTS free?

The open-source models are listed as free, and the project is licensed under Apache-2.0. DashScope API use has separate rates: $0.115 per 10,000 input characters for Voice Design and Voice Cloning, with output free; each lists a 110,000-character free quota.

Can I use Qwen3-TTS commercially?

Yes. The project lists commercial use as permitted and the repository uses the Apache-2.0 license.

What platforms does it support?

The listed platforms are API, Linux, self-hosted, and web. The project also provides a local Gradio web interface.

Can it clone a voice?

Yes. Voice cloning uses reference audio and its transcript; the README says using only the speaker embedding without a transcript may reduce cloning quality.

What models and output format are available?

Released models include 0.6B and 1.7B variants for custom voice or base tasks, plus a 1.7B voice design model. WAV is a listed export format.

Does it support online serving with vLLM-Omni?

The README says vLLM-Omni supports offline Qwen3-TTS inference. Online serving is planned for later support.

Facts from Qwen3-TTS's own pages

API access: Yes · github.com, 20 Sept 2026
API integration: The project links Alibaba Cloud DashScope real-time APIs for custom voice, voice cloning, and voice design. · github.com, 7 Oct 2026
API pricing: Alibaba Cloud lists international Qwen3-TTS API rates of $0.115 per 10,000 input characters for Voice Design and Voice Cloning, with output free; the page also lists a 110,000-character free quota for each. · alibabacloud.com, 7 Oct 2026
API requirements: The linked Alibaba Cloud real-time synthesis guide says API use requires configuring an API key and installing the latest DashScope SDK when using that SDK. · help.aliyun.com, 7 Oct 2026
Commercial use: Yes · github.com, 20 Sept 2026
Download and install: Models can be downloaded through Hugging Face or ModelScope, and the project provides a `qwen-tts` Python package installable from PyPI. · github.com, 7 Oct 2026
Export formats: WAV · github.com, 20 Sept 2026
Generation modes: The models support streaming and non-streaming speech generation, voice design, voice cloning, and natural-language voice control. · github.com, 7 Oct 2026
Hardware and setup: The README demonstrates model loading on CUDA and recommends FlashAttention 2 to reduce GPU memory use, subject to compatible hardware. · github.com, 7 Oct 2026
Inference integration: The README states that vLLM-Omni supports offline Qwen3-TTS inference and that online serving is planned for later support. · github.com, 7 Oct 2026
Languages: 10 · github.com, 20 Sept 2026
Languages: The project lists support for Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, as well as multiple dialectal voice profiles. · github.com, 7 Oct 2026
Latency: The README says the streaming architecture can produce the first audio packet after one character and reports end-to-end synthesis latency as low as 97 ms. · github.com, 7 Oct 2026
License: The GitHub repository lists the Apache-2.0 license. · github.com, 7 Oct 2026
Local installation: The project provides a Python package install and says it recommends a fresh Python 3.12 environment. · github.com, 7 Oct 2026
Local web interface: A local Gradio web UI demo can be launched with the qwen-tts-demo command. · github.com, 7 Oct 2026
Model hosting: The README links model downloads on Hugging Face and ModelScope and documents loading weights locally or by model ID. · github.com, 7 Oct 2026
Model sizes: Released models include 0.6B and 1.7B parameter variants for custom voice or base tasks, plus a 1.7B voice design model. · github.com, 7 Oct 2026
Platforms: web,api,self_hosted · github.com, 20 Sept 2026
Product: Qwen3-TTS is an open-source series of text-to-speech models developed by the Qwen team at Alibaba Cloud. · github.com, 7 Oct 2026
Serving limitation: The README says vLLM-Omni supports offline inference currently, with online serving planned for later. · github.com, 7 Oct 2026
Streaming: The README says the models support streaming and non-streaming generation and reports end-to-end synthesis latency as low as 97 ms. · github.com, 7 Oct 2026
Voice cloning: The Base models can clone a voice from a reference audio clip and are described as capable of rapid cloning from three seconds of input audio. · github.com, 7 Oct 2026
Voice cloning input: Voice cloning uses reference audio and its transcript; the README says using only the speaker embedding without a transcript may reduce cloning quality. · github.com, 7 Oct 2026