RAM is the single most common bottleneck we see when someone tries to run AI locally and concludes "my computer cannot do this." Usually the computer can — it just needs memory.
By Use Case
- 16GB — meets the Copilot+ minimum. Fine for built-in Windows AI features. Tight for anything you load yourself.
- 32GB — the realistic floor for running your own models alongside normal work. This is what we recommend to most people.
- 64GB — comfortable for larger models, multiple applications, and virtual machines at the same time.
- 128GB+ — specialist territory. Worth it for large models or serious data work, otherwise money better spent elsewhere.
Why VRAM Matters More Than System RAM
If you have a dedicated graphics card, the model ideally loads into the GPU's own VRAM, which is far faster than system memory. A 12GB card handles a 7B-class model comfortably. When the model does not fit in VRAM it spills into system RAM and slows down dramatically — that is the moment people assume something is broken.
Before You Buy More RAM
- Check how many slots your machine has and how many are populated.
- Check the maximum supported capacity for your specific motherboard, not the chipset.
- Match speed and timings to the existing kit, or replace it as a matched pair.
- On laptops, confirm the memory is not soldered — on many thin models it is.
MediaVision Technical Analysis
Adding RAM is one of the cheapest meaningful upgrades still available, and on a machine that already feels slow it usually helps more than a new CPU would. We check compatibility before ordering anything, which avoids the most common and most annoying failure: buying memory that the board will not post with.
