RepoPicks Radar

Updated every day around 23:30 UTC

Guides

Qwen3-TTS

QwenLM/Qwen3-TTSReviewed by AI

Checked by RepoPicks Radar on

What is Qwen3-TTS?

Alibaba's Qwen team released these speech models with a Python package. You can read text aloud in ten languages with nine ready voices, describe the voice you want in plain words and let it invent one, or give it a three-second recording and have it speak in that voice. You install Python and the qwen-tts package, and the model files (0.6B or 1.7B) download on first use. The examples all load the model on an NVIDIA card. One command opens a local web page for trying it without writing code; Alibaba also sells the same thing as an online service.

Qwen3-TTS is free and open source (Apache-2.0) and can replace ElevenLabs, Murf, and Fish Audio.

How easy is it to install Qwen3-TTS?

D · terminal, several tools or a specific graphics card · Windows

  • Windows D
  • Linux D

On Linux and Windows it is terminal work plus specific hardware: create a Python environment, "pip install -U qwen-tts", let a model of a few gigabytes download, then run "qwen-tts-demo <model>" for the local web page. Every example in the README loads the model with device_map="cuda:0" and bfloat16 and recommends building flash-attn, so an NVIDIA card is expected; Linux is the easier of the two because flash-attn and vLLM are built for it. Mac is not mentioned at all.

Coming soon: install in 1 click with the RepoPicks app.

Qwen3-TTS on GitHub

Stars
14,343on Nov 10, 2026
Downloads
—all releases
Growth
+128stars in 6 days

Stars over time

14,20014,25014,30014,350 Nov 5Nov 6Nov 7Nov 8Nov 9Nov 10 14,20014,25014,30014,350 Nov 5Nov 8

Facts

Runs on
Windows, Linux
You need to install
Python, NVIDIA graphics card
To build from source
Python
Disk space
150 MB–450 MB · estimate
Accounts or keys
None
Cost
Free · Code and model weights are free under Apache-2.0; Alibaba Cloud's DashScope offers the same models as a paid online service.
License
Apache-2.0 (Permissive)
Last change on GitHub
Mar 17, 2026

Qwen3-TTS risk signals (5)

  • Attentionvoice cloning of real people

    The Base models are described as doing "3-second rapid voice clone from user audio input", and ref_audio "can be a local file path, a URL, a base64 string", so three seconds of someone's voice taken from the internet is enough; copying a person's voice is restricted in several countries.

    AI review

  • Attentiondemo server opens to the network without a password

    The README's own commands are "qwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-Base --ip 0.0.0.0 --port 8000" and, for HTTPS, a self-signed certificate with "--no-ssl-verify"; 0.0.0.0 means anyone who can reach your machine on that port can use the page, which has no login.

    AI review

  • Infovoice design example describes a childlike voice

    One Voice Design example in the README asks for "撒娇稚嫩的萝莉女声" (a childish, babyish girl voice) and another for a "Male, 17 years old" voice, so the model can be told to produce voices that sound like minors.

    AI review

  • Infomodel files download on first use

    README: "During model loading in the qwen-tts package or vLLM, model weights will be automatically downloaded based on the model name", from Hugging Face or ModelScope; the 150-450 MB estimate in the checks covers the code, not those files.

    AI review

  • Infono releases and quiet since march

    There are no GitHub releases, the package version in pyproject.toml is 0.1.1, and the last commit is from 2026-03-17 (237 days before the check), while the README says other models from the report "will be released in the near future".

    AI review

Evidence is shown in its original language.

What Radar could not check about Qwen3-TTS

  • Whether it runs at all without an NVIDIA card, since the README shows no CPU or Apple Silicon path.
  • How much disk and video memory the 0.6B and 1.7B models really take after downloading.
  • The headline numbers (97 ms first audio, the error rates in the tables) were measured by the authors.
  • How good the Portuguese and Spanish voices are compared with Chinese and English, which get most of the attention.
  • Whether building flash-attn works on a normal home machine; the README warns about memory and CPU cores during that step.

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub