Buy memory,not speed.

Every machine on this page has a profile: its memory, how much of that memory a model can actually use, its bandwidth, and where it stands between the budget cards and the workstation class. From the profile and the measured file sizes the fit engine computes, for every tracked model, whether it runs comfortably, tightly, or not without offloading, at four reference contexts. The grid below shows the count of comfortable models per machine; each machine's own page shows every verdict and the arithmetic behind it.

The usable memory rule is the one assumption that decides most verdicts. A discrete card plans with a declared share of its sticker memory, because the driver, the display and the allocator keep the rest; a Mac plans with the share of unified memory macOS lets the GPU take, which is lower and stated on every Mac page. The engine never rounds in the machine's favour: a model that lands a few hundred megabytes over the line is rated as not fitting, because that is what happens on a desktop with a browser open.

The page says nothing about speed. Two machines with the same memory get the same verdicts here even when one of them produces tokens three times faster, because speed depends on the runtime, the quantization, the batch size and the driver, none of which this site measures. Memory decides whether a model runs at all; the hardware pages tell you that, and leave the rest to a benchmark you trust.