VRAM

VRAM is the graphics card's operating memory. When running AI locally, it's the main limitation: the model must fit in VRAM to work quickly. Hence the first question in self-hosting isn't "what processor," but "how much VRAM do you have." Roughly: small models fit in 8 GB, medium ones need around a dozen, and large ones require tens of gigabytes. Quantization lets you squeeze a larger model into less memory at a slight quality cost; an alternative is unified memory in Apple computers.