1,056-run local LLM benchmark on an M5 Max: six local models vs two frontier APIs through the pi agent harness - View it on GitHub