How to Setup a Local Coding Agent on macOS

A developer successfully set up a local coding agent on macOS using Gemma 4 26B model with llama.cpp and Metal acceleration, achieving 72.2 tokens/second—a 24% speedup—by adding a Multi-Token Prediction (MTP) draft model for speculative decoding. The setup runs on an Apple M1 Max and provides fast enough performance for practical coding work through an OpenAI-compatible API, outperforming alternative MLX implementations.