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

🔗 GitHub Source & Web Demo: vanch007/mlx-AuK · 8-Bit Quantized Base Weights · 4-Step AuK-Flash

This repository provides the native Apple Silicon MLX port of the full AuK Base foundation model, Tencent Hunyuan's 1.5B speech foundation model designed for high-fidelity speech generation, voice cloning, and zero-shot speech editing with configurable NFE steps and CFG guidance.

• 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: Configurable NFE (default 32 steps) & CFG strength (default 2.0)
• Sampling Rate: 24kHz Mono High-Fidelity Audio

Performance on Apple Silicon

Metric	PyTorch MPS	Native MLX (This Model)	Native MLX 8-Bit
32-Step Sampling Latency	29.8s	7.92s	8.15s
Real-Time Factor (RTF)	2.98	0.792 (1.26x real-time)	0.815
Backbone Memory	5.70 GB	5.70 GB	0.56 GB

Files Included

• : Full-precision MLX weights for the AuK Base Flux2Edit backbone (5.7GB).
• : Full-precision MLX weights for BigVGAN Flow VAE (608MB).
• : Architecture and sampling parameters.

Quickstart