I don’t know anything about training, but regarding inference: 

I just upgraded from 32gb m1 that I couldn’t run any worthwhile models with. Don’t consider anything under the m5 series because the prompt processing is far slower and you will wait substantially longer between each response

My new m5 pro 64 fits many of the popular models and can do 30B models in q8 with lots of context Dense models are still feeling slow and I can’t fit the 120B MoE models that I hear such good things about even at q4. When I want to talk to a dense model on the m5 pro it involves waiting for about 5 minutes between responses (~10 tps at low context).

An m5 max allows 128gb ram but the difference in price is pretty substantial. I wish I could have afforded the cost but double the memory bandwidth of the pro means the difference between running larger models at all, even though they’ll be quantized, the big MoEs will still feel fast. 27B class models can be run unquanted or at q8 and feel actuallyi useable for discussion.