We ran Qwen3 14B and Llama 3.2 3B on an Apple M2 with 24 GB unified memory. The 3B model is 3.5x faster, uses 4x less memory, and on a smal…