In general for local LLMs, the more memory the better. You will be able to fit larger models in RAM. The faster CPU will give you more tokens/second, but if you are just chatting with a human in the loop, most recent M series macs will be able to generate tokens faster than you can read them.
That also very much depends on model size. For 70B+ models, while the tok/s are still fast enough for realtime chat, it's not going to be generating faster than you can read it, even on Ultra with its insane memory bandwidth.
