Well, deepseek-r1:7b on AMD CPU only is ~12 token/s, gemma3:27b-it-qat is ~2.2 token/s. That's pure CPU at about 0.1x of a $3,500 Apple laptop at about 0.1x of the price. It's more a question about your patience, use case, and budget.

For discrete GPUs, RAM size is a harder cutoff. You either can run a model, or you can't.