---
base_model: Qwen/Qwen3-TTS-12Hz-0.6B-Base
library_name: mlx
tags:
  - mlx
  - tts
  - mxfp4
  - apple-silicon
  - qwen3-tts
  - microscaling
---

# Qwen3-TTS-12Hz-0.6B-Base — MXFP4 (MLX)

MXFP4 quantized version of [Qwen/Qwen3-TTS-12Hz-0.6B-Base](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-0.6B-Base) for Apple Silicon.

Converted using [mlx-audio](https://github.com/Blaizzy/mlx-audio) with native MXFP4 (Microscaling Float 4-bit, OCP MX Spec).

## Benchmark (M2 Ultra 128GB)

| Quant | Size | Avg Time (3 runs) |
|---|---|---|
| 8bit | 1.9 GB | 8.50s |
| **mxfp4** | **1.6 GB** | **7.77s (~8.6% faster)** |

Audio quality verified: voice cloning works, long German texts direct speech render cleanly.

## Conversion
```bash
python -m mlx_audio.convert \
  --hf-path Qwen/Qwen3-TTS-12Hz-0.6B-Base \
  --mlx-path ./Qwen3-TTS-0.6B-Base-mxfp4 \
  --quantize \
  --q-mode mxfp4
```

## Usage
```python
from mlx_audio.tts.utils import load_model
from mlx_audio.tts.generate import generate_audio

model = load_model("mpe74/Qwen3-TTS-12Hz-0.6B-Base-mxfp4")
generate_audio(
    model=model,
    text="Hello, this is a test.",
    ref_audio="reference.wav",
    temperature=0.3,
    repetition_penalty=1.1,
)
```

## CLI
```bash
python -m mlx_audio.tts.generate \
  --model mpe74/Qwen3-TTS-12Hz-0.6B-Base-mxfp4 \
  --text "Dies ist ein Test." \
  --ref_audio reference.wav \
  --ref_text "Transkript der Reference Audio" \
  --temperature 0.3 \
  --repetition_penalty 1.1 \
  --play
```