the new open weights model is huge by parameter count but it's mixture of experts, so only a fraction of it activates for any given token and a single box with enough memory can actually serve it for chat. the catch is that every expert still has to sit in memory, you can't load it piecemeal, and the smallest quant worth running is over a hundred gigabytes. so you need something like a hundred twenty eight gigs of memory and that's the binding constraint.
the three realistic machines all carry about that much. a mac notebook with fast unified memory, a tiny nvidia cuda box, and an amd mini pc that hit the memory target first. the numbers get weird here because decode on an active-fraction-of-billion model is memory bandwidth bound, so all three land around thirty tokens a second, and the mac's huge bandwidth advantage barely shows. roofline analysis explains it, the cuda box hits most of its bandwidth ceiling while the mac converts only a fraction of its theoretical bandwidth into tokens.
the bigger trap is quality. one box of this memory class can only hold the compressed quant, and at release there was no validated build of that quant for this model. for a coding harness, raw speed is worthless if the tokens start lying, and coding reliability degrades before casual chat makes the damage obvious. the reliable path is either two of the little cuda boxes so you can hold a proper quant, or accepting the lossy build for casual chat only.
so by price per prefill token the tiny cuda box wins, but the honest answer is two of them for anything you actually trust with code. a single box at full speed runs the wrong model, double the money runs you the right model at the same real speed. decide which one you're buying before you buy anything.