Bringing Voice AI to Apple Silicon: How to Run LLMs Locally on a Memory-Constrained Mac

This project demonstrates how to run multiple ML models simultaneously on Apple Silicon with just 8GB of unified memory by carefully selecting lightweight alternatives like MLX-optimized Whisper and Qwen 2.5. The shift from server-side to browser-based audio encoding solves a critical portability bottleneck, making local voice AI accessible without requiring users to install external dependencies. This approach highlights emerging best practices for deploying generative AI on consumer hardware while maintaining privacy and eliminating cloud dependencies.