🚀 AuK-Flash-MLX (Native Apple Silicon Port)

🔗 GitHub Source & Web Demo: vanch007/mlx-AuK · 8-Bit Quantized MLX Weights

This repository provides the native Apple Silicon 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
• Architecture: 10 Double-Stream MMDiT Blocks + 20 Single-Stream DiT Blocks + Causal BigVGAN-Flow-VAE + Qwen2.5-Omni Thinker
• Inference Mode: 4-Step Distilled Flow Matching (DMD) with fixed CFG=0.0
• Sampling Rate: 24kHz Mono High-Fidelity Audio

Performance on Apple Silicon

Metric	PyTorch MPS	Native MLX (This Model)	MLX 8-Bit Quantized
4-Step DiT Latency (10s Audio)	3.820s	0.992s (3.85x faster)	1.022s
Real-Time Factor (RTF)	0.3820	0.0992 (10.08x real-time)	0.1022 (9.79x real-time)
Backbone Memory	5.70 GB	5.70 GB	0.56 GB (90.1% saving)

Files Included

• : Full-precision MLX weights for the Flux2Edit diffusion transformer backbone (5.7GB).
• : Full-precision MLX weights for the BigVGAN Flow VAE (608MB).
• : Architecture and model hyperparameters.

Quickstart & Interactive Demo

For complete usage instructions, Python API, and the interactive A/B comparison web demo, visit: 👉 https://github.com/vanch007/mlx-AuK