VRAM

For a discrete graphics card, dedicated memory is separate from ordinary system RAM. Some computers use a unified memory arrangement instead, so interpret the hardware report for that machine. A GPU model name alone is not a usable capacity check.

Model weights are only part of the allocation. The runtime needs working buffers, and language-model attention caches can become substantial for long contexts. Hugging Face explains cache memory trade-offs.

Leave room for the display and other running applications. If a runtime moves parts of a model into system memory, it may fit but behave differently from an entirely GPU-resident run. Check actual allocation and supported hardware instead of relying on a file-size comparison. Ollama's FAQ explains its CPU/GPU placement display.