On this page
You can run Qwen3-TTS offline on an Apple silicon Mac with mlx-audio, a free, open-source Python package built on Apple’s MLX. Install it with one pip command, run one line in Terminal, and Qwen3-TTS speaks on your Mac with no cloud service. After the first download, it works with Wi-Fi off.
Off Grid AI doesn’t run Qwen3-TTS today. If you want offline voice without a terminal, Off Grid AI Desktop reads chat replies aloud with its built-in Kokoro voice. That’s covered at the end.
Download Off Grid AI Desktop (free, no account) for offline voice in a chat app.
What Qwen3-TTS can do
Qwen3-TTS is Alibaba’s open text-to-speech family, under the Apache 2.0 licence. It speaks ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. It comes as five models that do three jobs.
| Model | What it does | MLX download (8-bit) |
|---|---|---|
| 0.6B CustomVoice | Nine preset voices | about 2.0 GB |
| 1.7B CustomVoice | Nine preset voices, plus tone and emotion from an instruction | about 3.1 GB |
| 0.6B Base | Clones a voice from a short clip | about 2.0 GB |
| 1.7B Base | Clones a voice, higher quality | about 3.1 GB |
| 1.7B VoiceDesign | Makes a new voice from a description | about 4.5 GB (bf16) |
Sizes are the mlx-community files plus the 0.68 GB speech tokenizer each one needs. Qwen says the Base models clone a voice from about 3 seconds of audio, and that streaming starts in as little as 97 ms. Those are Qwen’s figures, not ours.
What you need
| Mac | Apple silicon (M1 or later). MLX doesn’t run on Intel Macs |
| Memory | 8 GB is enough. mlx-audio measured 3.9 GB for the 1.7B CustomVoice model in 6-bit |
| Python | 3.10 or later. The Python that ships with macOS is older, so install a current one from python.org or Homebrew |
| Disk | 2 to 5 GB per model |
| Cost | Free |
How to set it up
Open Terminal and check your Python first.
python3 --version
If it says 3.9 or lower, install a newer Python and open a new Terminal window. Then make a folder for it and install mlx-audio:
python3 -m venv ~/qwen3-tts
source ~/qwen3-tts/bin/activate
pip install -U mlx-audio
Now make Qwen3-TTS speak. This uses the small preset-voice model and plays the result:
mlx_audio.tts.generate \
--model mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit \
--text 'Hello! This voice is running on my Mac.' \
--voice Ryan --lang_code English --play
The first run downloads about 2 GB from Hugging Face. Every run after that uses the copy on your Mac. To save the audio instead of playing it, add --output_path ./my_audio.
Pick a voice
The CustomVoice models have nine voices. Each sounds best in its own language, though any of them can speak all ten.
| Voice | Sounds like | Native language |
|---|---|---|
| Ryan | Energetic male | English |
| Aiden | Clear American male | English |
| Vivian | Bright young female | Chinese |
| Serena | Warm, gentle young female | Chinese |
| Uncle_Fu | Low, mellow older male | Chinese |
| Dylan | Young Beijing male | Chinese (Beijing dialect) |
| Eric | Lively Chengdu male | Chinese (Sichuan dialect) |
| Ono_Anna | Playful female | Japanese |
| Sohee | Warm, expressive female | Korean |
Voice descriptions are summarised from Qwen’s model card.
Clone a voice
Cloning uses a Base model. Record 5 to 10 seconds of clear speech as a WAV file, then pass the file and the exact words spoken in it. Don’t add --voice when you clone, because mlx-audio says the two settings conflict.
mlx_audio.tts.generate \
--model mlx-community/Qwen3-TTS-12Hz-0.6B-Base-8bit \
--text 'This is my cloned voice reading a new sentence.' \
--ref_audio ./my_voice.wav \
--ref_text 'The exact words I said in the recording.' \
--play
Only clone your own voice, or a voice you have permission to use.
Design a voice from a description
The 1.7B VoiceDesign model makes a voice from a sentence describing it:
mlx_audio.tts.generate \
--model mlx-community/Qwen3-TTS-12Hz-1.7B-VoiceDesign-bf16 \
--text 'Welcome back. Here is your summary for today.' \
--instruct 'A calm, low male voice with a slow, steady pace.' \
--lang_code English --play
Use it fully offline
Once a model has downloaded, turn off Wi-Fi and run the same command. To stop the tools from trying the network at all, start each Terminal session with:
export HF_HUB_OFFLINE=1
Your text never leaves your Mac, because the model runs on it.
If something goes wrong
| Symptom | Check | Fix |
|---|---|---|
pip fails or mlx-audio won’t install |
python3 --version |
Install Python 3.10 or later, then redo the venv |
command not found: mlx_audio.tts.generate |
Is the venv active? | Run source ~/qwen3-tts/bin/activate |
| It fails on an Intel Mac | Your chip | MLX needs Apple silicon |
| Cloning errors out | Did you pass --voice? |
Remove --voice when you use --ref_audio |
| The clone sounds off | Your recording | Use a quiet, clear clip and its exact transcript |
| An accent sounds wrong | The voice’s native language | Pick a voice whose native language matches your text |
| It tries to reach the internet | Model cache | Run it once online, then set HF_HUB_OFFLINE=1 |
Offline voice without a terminal
If you want your Mac to read replies aloud and never touch Terminal, Off Grid AI Desktop has a local voice built in. It uses Kokoro, not Qwen3-TTS.
- Download Off Grid AI Desktop and open it.
- Open Settings > Voice. Choose a Language, then a Voice, and wait for the voice to be ready.
- Select Test voice to hear it.
- Chat with a local model, then use Speak on any reply.
The local voice covers US and UK English, French, Spanish, Italian, Portuguese, Hindi, Polish and German. It runs on your Mac and keeps working offline after setup.
FAQ
Can Qwen3-TTS run on a Mac? Yes, on Apple silicon. The open-source mlx-audio package runs Qwen3-TTS on Apple’s MLX framework. Install it with pip, and the first run downloads the model from Hugging Face. After that it works offline.
Does Off Grid AI run Qwen3-TTS? Not today. Off Grid AI Desktop speaks replies offline with its built-in Kokoro voice, in Settings > Voice. Use mlx-audio if you need Qwen3-TTS itself.
Which Qwen3-TTS model should I start with? Qwen3-TTS 0.6B CustomVoice in 8-bit. It’s about 2 GB with its speech tokenizer and has nine preset voices. Move to a 1.7B model for better emotion control or to design a voice from a description.
Can Qwen3-TTS clone my voice on a Mac? Yes, with a Base model. Give it a short recording and the exact words spoken in it. Qwen says the Base models clone from about 3 seconds of audio. Only clone your own voice or one you have permission to use.
What languages does Qwen3-TTS speak? Ten: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.
Related guides
- How to Turn Text Into Speech in Multiple Languages on Your Mac in 2026
- How to Run Text-to-Speech Locally on Your Desktop in 2026 (Kokoro)
- How to Run Qwen3-ASR on Your iPhone in 2026
Download Off Grid AI Desktop for offline voice on your Mac with no setup.