| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2025-10-01 | 2.1 kB | |
| v1.8.0 - Quantized 8-Bit Model _ Dynamic 8-Bit Quantization source code.tar.gz | 2025-10-01 | 100.6 kB | |
| v1.8.0 - Quantized 8-Bit Model _ Dynamic 8-Bit Quantization source code.zip | 2025-10-01 | 125.0 kB | |
| Totals: 3 Items | 227.6 kB | 0 | |
🎉 Quantized 8-Bit Model Release
- New Model: VibeVoice-Large-Q8 - Perfect quality with 40% less VRAM
- Dynamic 8-Bit: Added 8-bit option for on-the-fly quantization
📦 VibeVoice-Large-Q8
Download
- HuggingFace: FabioSarracino/VibeVoice-Large-Q8
- Size: 11.6GB (38% smaller than full precision)
- VRAM: ~12GB (40% reduction from ~20GB)
Key Innovation
- Perfect Audio Quality: Identical to full precision - zero degradation
- Selective Quantization: Audio-critical components remain at full precision
- Diffusion head ✓ Full precision
- VAE ✓ Full precision
- Connectors ✓ Full precision
- LLM ✓ 8-bit quantized
🔧 Dynamic 8-Bit Quantization
Enhanced quantize_llm Parameter
Now supports three options:
- full precision (default) - Original quality
- 8bit (NEW) - Balanced quality/VRAM
- 4bit - Maximum VRAM savings
Smart Component Selection
8-bit mode intelligently skips:
- Diffusion head
- Acoustic/semantic connectors
- Tokenizers
- All audio-processing components
💾 VRAM Comparison
| Model/Mode | VRAM Usage | Quality |
|---|---|---|
| VibeVoice-Large | ~20GB | Perfect |
| VibeVoice-Large-Q8 | ~12GB | Perfect |
| Dynamic 8-bit | ~13-14GB | Excellent |
| Dynamic 4-bit | ~10-11GB | Very Good |
| VibeVoice-Large-Q4 | ~8GB | Good |
🎯 Perfect For
- RTX 3060 (12GB): Run VibeVoice-Large-Q8 with perfect quality
- RTX 4070 Ti (12GB): Full model without compromises
- RTX 3050 (8GB): Use VibeVoice-Large-Q4 or dynamic 4-bit
- Production: Best quality-to-VRAM ratio with Q8
- Development: Fast iteration with dynamic quantization
⚙️ Requirements
For dynamic quantization: - NVIDIA CUDA GPU - bitsandbytes library - Falls back to full precision on CPU/MPS
💾 Installation
Install via ComfyUI Manager or manually:
:::bash
git clone https://github.com/Enemyx-net/VibeVoice-ComfyUI