This model contains the original weights plus a quantization_config in config.json. When you load the model, it will automatically be quantized to 4-bit using bitsandbytes NF4 quantization.

Memory savings: ~75% reduction compared to full precision.

Requirements

pip install qwen-tts bitsandbytes>=0.42.0 accelerate

Usage

python

from qwen_tts import Qwen3TTSModel
import torch

# Model will be automatically quantized when loaded
model = Qwen3TTSModel.from_pretrained(
    "YOUR_USERNAME/Qwen3-TTS-12Hz-1.7B-Base-BNB-4bit",
    device_map="auto",
)

# Voice cloning example
import soundfile as sf

wavs, sr = model.generate_voice_clone(
    text="Hello, this is a test of the quantized model.",
    language="English",
    ref_audio="path/to/reference.wav",
    ref_text="Transcript of your reference audio.",
)
sf.write("output.wav", wavs[0], sr)

Quantization Details

Setting	Value
Method	bitsandbytes
Bits	4
Quant Type	NF4 (Normalized Float 4)
Compute Dtype	bfloat16
Double Quant	Yes

Original Model

Based on Qwen/Qwen3-TTS-12Hz-1.7B-Base. Please refer to the original model card for full documentation.

License

Apache 2.0 (same as original model)

divyajot5005/Qwen3-TTS-12Hz-1.7B-Base-BNB-4bit

Qwen3-TTS-12Hz-1.7B-Base - 4-bit Quantized (bitsandbytes)

How It Works

Requirements

Usage

Quantization Details

Original Model

License

divyajot5005/
Qwen3-TTS-12Hz-1.7B-Base-BNB-4bit