VRAM
VRAM is the dedicated memory on your graphics card that holds model weights and key tensors during local LLM inference.
VRAM is not system RAM. If a 70B model needs 40 GB of weights at a given quant, a 12 GB consumer GPU cannot load it fully on-card without offloading layers to CPU (which slows tokens per second).
Planning VRAM means adding headroom for the context window and any vision or tool-use adapters. WikiWayne VRAM matrix charts on pillar articles summarize common model sizes against 8, 12, 16, and 24 GB tiers.