Because OP is running it on an M3 Ultra with Ollama. He'll get much better perf with llama.cpp or MLX, both of which Ollama wraps, albeit very poorly.