Is Qwen3-TTS open source?

Yes. Qwen3-TTS is an open-source model from Alibaba's Qwen team, and the code and weights are available from the official repository at github.com/QwenLM/Qwen3-TTS, where you can check the current license terms.

How much VRAM does Qwen3-TTS need?

Around 8 GB of video memory is enough to run Qwen3-TTS comfortably. It is lighter than its quality suggests, so a mainstream card works. The manual route needs NVIDIA; the desktop app also runs on AMD and Intel Arc cards, Apple Silicon Macs, or the CPU alone, more slowly. You can also run it in Voice Creator Pro Cloud from any browser, with no local GPU involved.

Do I need flash-attention to run Qwen3-TTS?

No. flash-attention is an optional speedup, and it often fails to build on Windows. Qwen3-TTS runs without it. In the Voice Creator Pro desktop app you don't have to deal with it, because the app ships with everything already set up and there is nothing for you to build.

What emotions can Qwen3-TTS do?

In Voice Creator Pro, Qwen3-TTS offers 13 emotions, each at 5 intensities, and they work on cloned voices as well as preset and designed ones. The open-source model on its own takes emotion as a written instruction rather than a set menu.

Is Qwen3-TTS good for audiobooks?

Yes. Consistent clones, clean handling of numbers and abbreviations, and emotion control make it a strong narration model. It is slower than lighter models, so for book-length projects the desktop app's file import and unlimited generations help, or you can let Voice Creator Pro Cloud do the processing.

Can I run Qwen3-TTS on a Mac?

Yes, with the Voice Creator Pro desktop app, which runs Qwen3-TTS locally on Apple Silicon Macs (M1 or newer). The manual open-source route is written for NVIDIA GPUs, so on a Mac you would be working out a different setup on your own. Voice Creator Pro Cloud also works in any browser.

Why is Qwen3-TTS so slow?

Qwen3-TTS trades speed for expressiveness, so it generates more slowly than lighter models like OmniVoice even on a good GPU, and it slows to a crawl if it falls back to the CPU. Voice Creator Pro Cloud runs it on server GPUs, which takes the load off your machine.

How to Run Qwen3-TTS Locally: The Easy Way and the Manual Way (2026)

Qwen3-TTS is the open-source model for scripts where the feeling matters as much as the words. Built by Alibaba's Qwen team, it clones a voice from a short sample, designs a new voice from a plain-language description, reads numbers and abbreviations cleanly, and produces some of the most consistent clones of any open model. However, it has a reputation for being fiddly to install.

That reputation is why this guide starts with the shortcut. If you want Qwen3-TTS working on your own computer today, the first section gets you there in three steps. After that we cover why manual installs break, and then sketch the do-it-yourself route for people who want to run the raw model.

Three ways to run Qwen3-TTS

	Install it yourself	Voice Creator Pro desktop app	Voice Creator Pro Cloud
What you set up	Python, PyTorch, GPU libraries, pinned `transformers` versions, and the model, all worked out yourself	Just the app; it packages everything Qwen3-TTS needs	Nothing; it runs in your browser
When something breaks	You fix it yourself, often after a library update	All dependencies included and tested; support helps if anything breaks	Nothing on your computer to break; support helps if needed
Hardware	NVIDIA GPU only (~8 GB)	NVIDIA, AMD, Intel Arc, or Apple Silicon; CPU-only works too, slower	Any browser
Interface	Command line and Python scripts	Desktop app	Browser app
Emotion	Written instructions; none on cloned voices	13 emotions at 5 intensities	13 emotions at 5 intensities
Long documents	You split and feed the text yourself	Import an EPUB, PDF, Word doc, or text file	Import an EPUB, PDF, Word doc, or text file
Runs offline	Yes, once set up	Yes	No
Cost	Free model; you supply the hardware (check the license)	One-time purchase, unlimited generations (see prices)	Free to generate and listen; downloads from $5 a month

The easy way: Qwen3-TTS in the Voice Creator Pro desktop app

Voice Creator Pro is a desktop app for Windows and Apple Silicon Macs with Qwen3-TTS built in. Qwen3-TTS runs on Python, PyTorch, the GPU libraries for your graphics card, and specific versions of Hugging Face's transformers. The app packages all of them for you, so you never have to find the right versions, install them, or fix them when an update breaks something.

It runs on NVIDIA, AMD, or Intel Arc graphics cards (8 GB of video memory recommended), on Apple Silicon Macs, or on the CPU alone.

Install the app from the Microsoft Store (free trial available) or the Mac App Store.
Choose Qwen3-TTS in the model list.
Generate or clone. Paste your script, pick a voice or clone one from a clean 3 to 10 second clip, set an emotion, and generate.

The desktop app runs Qwen3-TTS locally and offline, with unlimited generations and nothing leaving your machine. It is a one-time purchase (see current prices) with no subscription.

What you get on top of the base model:

13 emotions at 5 intensities. Pick an emotion for each part of your script and choose how strongly it lands. This is a Voice Creator Pro layer: the base model does not offer an emotion menu, and it works on cloned voices too, so a cloned voice can be directed rather than read flat.
Free-form voice design. Describe the voice you want in your own words. The voice prompting guide shows what to put in a description.
Whole files, not snippets. Import an EPUB, PDF, Word doc, or text file and it narrates the lot, with no manual chunking, which suits Qwen3-TTS's audiobook strengths.
The rest of the toolkit. Other models (including OmniVoice for 600+ languages), video dubbing, subtitles, speech to text, and a voice changer, all in one app.

No install at all: Voice Creator Pro Cloud

Voice Creator Pro Cloud runs Qwen3-TTS in any browser on server GPUs, which also sidesteps the model's slow generation on modest hardware. The free tier lets you generate and listen with no card required. Downloading audio for publishing starts at $5 a month. Your text is processed on servers and never used to train models.

Common problems when installing Qwen3-TTS yourself

Qwen3-TTS has a few failure points of its own, on top of the usual GPU headaches.

Library version breakage. Qwen3-TTS loads through Hugging Face's transformers library, and major releases of that library have changed how models load. An install that worked last month can fail after a routine update, with an error that points deep inside the library rather than at the version mismatch. Staying working means pinning versions and watching the official repository for changes.

flash-attention will not build. Many guides tell you to install flash-attention for speed. It is optional, but on Windows it usually has to be compiled, which needs a matching compiler toolchain and CUDA setup, and the build often fails after a long wait. People assume the whole install is broken when only this optional extra is.

CUDA and driver mismatches. As with most GPU models, the PyTorch build has to line up with a CUDA version your NVIDIA driver supports. A plain install often grabs a CPU-only PyTorch, and a mismatched build throws CUDA errors at load time.

Conflicts inside ComfyUI and other tools. Running Qwen3-TTS through a ComfyUI node or alongside other AI tools means sharing one Python environment. The package or one of its dependencies can fail to install cleanly into that environment, leaving an import error that names the symptom but not the cause.

Running out of video memory. Qwen3-TTS is lighter than its quality suggests, around 8 GB of video memory, but on smaller cards or with other apps using the GPU it can still run out mid-generation.

Slow generation, or a crawl on CPU. Qwen3-TTS is the slowest of the popular open models even on a good GPU. If the GPU is not detected and it falls back to the processor, short clips can take many minutes, which makes it impractical for anything longer than a test.

The manual route, in brief

If you want to run the open-source release yourself, you will be working out what to install, where to get it, and which versions work together, then troubleshooting the problems above when they appear. Here is the outline. The exact commands and supported versions change between releases, so take them from the official repository rather than from any third-party guide, this one included.

Hardware. A Windows or Linux PC with an NVIDIA GPU with around 8 GB of video memory, 16 GB of system RAM, and several gigabytes of disk space for the weights. The model comes in more than one size, and the smaller one is the practical pick on modest cards. There is no comfortable CPU-only option.

The general shape of the process:

Create a separate Python environment so Qwen3-TTS's library versions do not collide with other tools.
Install a GPU-enabled PyTorch build that matches your driver's CUDA version.
Install the Qwen3-TTS package with the library versions the repository specifies, and skip optional speedups like flash-attention unless your setup can build them.
Download the model weights for the variant you want (preset voices, cloning, or voice design) and generate from a Python script.

Where to get the exact steps: the official repository from the Qwen team, Qwen3-TTS on GitHub. It lists current requirements, model variants, and license terms.

Running the base model this way, emotion is steered with written instructions rather than chosen from a menu, and there is no emotion control for cloned voices.

Who should run Qwen3-TTS themselves

Running it yourself makes sense if you are comfortable managing Python environments and pinning library versions, and you want to build Qwen3-TTS into your own code. The manual route needs an NVIDIA card with around 8 GB of video memory.

The desktop app or Cloud is the better fit if you have a Mac, an AMD or Intel Arc card, or no graphics card at all, you want to pick emotions from a menu (including on cloned voices), you are narrating long documents, or you do not want an update to break your setup.

If your priority is language coverage rather than emotion, OmniVoice speaks 600+ languages and is lighter to run; see How to Run OmniVoice Locally. For a side-by-side on the same cloned voice, read Qwen3 vs OmniVoice vs Chatterbox, and for the full feature rundown, visit the Qwen3-TTS model page.