Why Your 8GB VRAM GPU Shows 24GB in Windows — And What Actually Fits

A developer testing an RTX 5060 laptop found that Windows reports 24GB of display memory, but only 7.9GB is dedicated VRAM, with the rest being slower shared system RAM. The two memory pools behave very differently under load, making the advertised figure misleading for anyone trying to run AI models locally. Benchmarks on the same machine showed that dense models must be around 7GB or smaller at 4-bit quantisation to run reliably, while mixture-of-experts models can be larger if experts are streamed from system RAM. A model that had been running successfully for over a week suddenly failed to load, traced not to a bug but to a silently recreated local alias that altered memory calculations. The findings highlight that model architecture matters more than parameter count, and that available VRAM at runtime is narrower than any spec sheet suggests.