According to user benchmarks collected by a llama.cpp developer (7B model, 4-bit quantization), the M1 runs at about 14 tokens per second and the M2 or M3 at about 21 tokens per second.