Developer benchmarks 37 local LLMs on M5 MacBook Air for open-source performance database

The benchmark highlights the practical limits of running dense models on 32GB Macs, where 32B parameter models hit a wall at roughly 2.5 tokens per second. MoE (Mixture of Experts) models are proving essential for local inference on consumer hardware, offering significantly higher speeds with similar intelligence to dense models. The 32GB RAM ceiling is a significant bottleneck for interactive chat with large dense models, making MoE the current most viable solution for advanced local AI. Community-driven benchmarks like this are increasingly valuable as the landscape of local models and Apple Silicon configurations expands rapidly.