Qwen3-TTS GGUF

Qwen3-TTS GGUF

Speaks text in several languages with named speakers and Mandarin dialects, and can clone a voice from a short clip, through the qwentts.cpp C++ runtime on CPU, CUDA, Metal or Vulkan. Suits developers who want local text to speech.

Made by Serveurperso on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3-TTS GGUF holds GGUF weights for qwentts.cpp, a C++ port of the Qwen3-TTS 12 Hz model from the Qwen team. It does zero-shot text to speech in several languages, with named speakers and Mandarin dialects, and outputs 24 kHz mono audio. It runs on CPU, CUDA, Metal and Vulkan.

Who it suits: Developers who want multilingual speech generation on their own machine through a C++ runtime, under Apache 2.0 for the model and MIT for the tooling.

Industry | Who uses it | How they use it | Example
Software | App developers | Add local multilingual speech to a product | Read text aloud in French
Media | Audio producers | Narrate scripts with a named speaker | Voice a short explainer
Global services | Localization teams | Speak in Chinese, English and six other languages | Make a Japanese audio prompt
Voice products | Agent builders | Build a custom voice from a clip you have the right to use | Clone your own voice for an assistant

Skip it if you want a Python-only package, since the card's quick start builds a C++ project.

Where to use it: It runs through qwentts.cpp on many kinds of hardware.

• CPU, with an x86 variant picked automatically
• NVIDIA GPUs through CUDA, the fastest path on Ada and Blackwell cards
• Vulkan GPUs from AMD, Intel or NVIDIA
• Apple Silicon GPUs through Metal

Use case | Who | What they get back
Named-speaker speech | Developers | A wav file in a chosen speaker's voice
Voice cloning | Creators | A wav file in the voice of a reference clip
Voice design | Designers | Speech from a described voice

How to set it up: You need a C++ build setup and the right model files.

• The qwentts.cpp project from the author's GitHub, cloned with its submodules
• A build for your hardware, such as the CUDA build script
• Two GGUF files that load together, a talker and a tokenizer
• Python with pip for the hub tools

1. Clone the qwentts.cpp project with its submodules and build it for your hardware.
2. Make a models folder.
3. Download a talker file and the tokenizer file into it.
4. Run one of the example scripts.

pip install -U huggingface_hub
hf download Serveurperso/Qwen3-TTS-GGUF

Once the project is cloned, the card's quick start builds it with the CUDA build script, makes a models folder, and downloads the Q8_0 talker and tokenizer files into it. Then it moves into the examples folder.

cd examples

How to use it: Run the example scripts from the examples folder. Each one writes a wav file, and you set GGML_BACKEND if you want to force a device.

./base.sh
./clone.sh
./customvoice.sh
./tts.sh

Tool | What it does
base.sh | Speaks with a named speaker and writes base.wav
clone.sh | Clones a voice from a clip and writes clone.wav
customvoice.sh | Runs custom voice mode and writes customvoice.wav
tts.sh | Runs voice design and writes tts.wav
GGML_BACKEND | Forces CUDA0, Vulkan0, Metal or CPU

Example: You run the base script after the download. The talker turns the text into 12 Hz codes, the tokenizer turns those into 24 kHz audio, and a wav file with a named speaker's voice appears.

For newcomers and experts: If you are new to this, follow the quick start with the Q8_0 files, which the card names as the recommended default. Run the base script first, and try cloning only after that works. Use only voices you have the right to use.

If you know the field, the talker file holds a Qwen3 language model, a code predictor head and an optional speaker encoder. The tokenizer file turns 12 Hz codes into audio. The tokenizer file is not quantized evenly, so codebooks and small tensors stay in F32 and some convolution kernels fall back to F16.

Clone and build qwentts.cpp
Download talker and tokenizer GGUF files
Select a backend
Talker makes 12 Hz codes from text
Tokenizer decodes them to 24 kHz audio

Good to know: The card offers base, customvoice and voicedesign modes, and voicedesign exists only for the 1.7B talker. The tokenizer is shared across every talker.

Title: Talker 1.7B file size by variant
F32: 7.7 GB
BF16: 3.9 GB
Q8_0: 2.1 GB
Q4_K_M: 1.2 GB

• License: Apache 2.0 for the Qwen3-TTS model and its codec, and MIT for qwentts.cpp
• Cost to run: none named, apart from your own hardware
• Data it reaches: the text you give it, any reference clip you supply and the model files on your computer
• Not for: cloning a voice without the speaker's permission

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: gguf
• Parameters: 915M
• Weights: about 457 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,347,066
• Likes: 67
• License: apache-2.0
• Languages: zh, en, fr, de, es, it, pt, ja
• Based on: Qwen/Qwen3-TTS-Tokenizer-12Hz
• Tags: gguf, text-to-speech, voice-cloning, voice-design, mandarin-dialects, ggml, qwen, qwen3, qwen3-tts

Links

• Model: https://huggingface.co/Serveurperso/Qwen3-TTS-GGUF
• Files: https://huggingface.co/Serveurperso/Qwen3-TTS-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Serveurperso/Qwen3-TTS-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.