🚀 AuK-Flash-MLX-8bit (Apple Silicon 8-Bit Quantized)
🔗 GitHub Source & Web Demo: vanch007/mlx-AuK · Full Precision (FP32/BF16) MLX Weights
This repository provides the 8-Bit affine quantized MLX port of AuK-Flash, Tencent Hunyuan's 1.5B speech foundation model for ultra-fast generation and zero-shot speech/lyric editing on Apple Silicon.
- Main GitHub Repository: https://github.com/vanch007/mlx-AuK
- Original Project: Tencent-Hunyuan/AuK
- Quantization: 8-bit group-wise affine quantization ()
- Weight Size: 0.56 GB (90.1% reduction from 5.70 GB)
- Inference Speed: 1.022s for 10.0s audio on Apple Silicon (RTF 0.1022, ~10x real-time)
Performance Benchmark
| Metric | PyTorch MPS | Native MLX FP32 | Native MLX 8-Bit (This Model) |
|---|---|---|---|
| 4-Step DiT Latency (10s Audio) | 3.820s | 0.992s | 1.022s |
| Real-Time Factor (RTF) | 0.3820 | 0.0992 | 0.1022 (9.79x real-time) |
| Backbone Memory | 5.70 GB | 5.70 GB | 0.56 GB (90.1% saving) |
Files Included
- : 8-bit quantized MLX weights for the Flux2Edit diffusion transformer (569MB).
- : MLX weights for the BigVGAN Flow VAE (608MB).
- : Architecture and quantization hyperparameters.
Quickstart
For full inference workflows, CLI tools, and the interactive A/B comparison web demo, visit the project GitHub: 👉 https://github.com/vanch007/mlx-AuK