VRAM is memory attached to a GPU
VRAM is the memory a graphics processor uses for data it needs close at hand. When a local language model runs on a discrete GPU, some or all of its weights and working buffers can be placed in that memory.
A GPU's advertised VRAM is a capacity number, not a promise that every byte is available to a model. The display, drivers, other applications, the runtime, and the operating system can all consume part of it.
Why model weights need memory
Model weights are the learned numbers that make the model behave as it does. More parameters generally mean more weight data. Quantization stores those numbers using fewer bits, which can make a model practical on a smaller GPU.
The downloaded model file is only one part of the runtime picture. Loading and using a model also needs runtime buffers and memory for the active context, so a file that appears to fit exactly may still fail to load.
How LLMGauge uses VRAM
LLMGauge compares an approximate model requirement with the dedicated VRAM entered in the hardware profile. It leaves room for named runtime and safety assumptions instead of treating the VRAM number as fully available.
The result is planning guidance. Actual free memory depends on your runtime, drivers, display session, context length, and what else is running on the computer.