M5 Max 128GB makes local AI practical

Hot take: 128GB is no longer a joke for local LLMs on a Mac, but it is still a capacity play first and a speed play second.

• –Apple officially supports 128GB on M5 Max, and the higher memory bandwidth should help the prefill-heavy part of long-context inference that used to feel painfully slow on older Apple Silicon.
• –Early community benchmarks are showing clear M5 Max improvements over M4 Max, especially in prompt processing, which is the bottleneck that matters most for agentic coding workflows.
• –Loading bigger models and larger context windows is now realistic, but decode speed still depends heavily on the model, quantization, and backend, so it will not feel like a desktop GPU rig.
• –For serious local coding agents, 128GB makes sense if you want 70B-120B-class models and long context; if you mostly run 7B-32B models, 64GB is probably the better value.
• –The real “sweet spot” has shifted from “can it run at all?” to “how much model and context do you actually need?”