What do you need to run the model you actually want?

Tell us the model you want to run and what you own. You get a ranked set of routes from here to there — including the ones that cost nothing, which win whenever they are correct.

How the recommendation is made

We size the model the way every other page here does: total parameters at your quantization, plus the KV cache for your context length, plus framework overhead. Mixture-of-Experts models are sized by TOTAL parameters, because every expert must be resident even though only some are read per token.

We compare that against what your machine can actually offer — a card's usable VRAM, or an Apple or mini-PC unified pool minus the share the OS keeps.

Then we rank every route that closes the gap. A route that costs nothing outranks one that costs money whenever it works: if a quantization at or above Q4_K_M fits, or a smaller model in the same family fits, or your card plus system RAM holds it at a usable speed, we recommend buying nothing.

What we check before recommending a part

• Power supply. A card needs roughly 30% headroom over its board power plus the rest of the system. Tell us your PSU and, if it is too small, we exclude that card and say why rather than recommending something that would trip your machine.
• Memory generation. DDR4 and DDR5 are keyed differently and will not fit each other's slots. We never offer a kit your board cannot take.
• Physical space. Slot width for a card; free DIMM slots for a kit — four modules need four slots, not two.
• Whether the upgrade is possible at all. Apple Silicon memory is fixed at purchase and laptop graphics chips are soldered. Where we cannot tell whether your laptop's memory is soldered, you get a caveat telling you to check — not a silent guess either way.
• Whether more RAM would actually help. A memory kit never adds VRAM — it buys offload across PCIe, which is far slower. If the result would run below about 5 tokens a second, we exclude that path and tell you the rate rather than selling you a kit that makes the model load and nothing more.

Why there are no prices here

You will not see a price next to a shop link here. Amazon's terms only permit showing a price beside a link into their store when it is fetched live from their product API, which we do not have — so rather than a number that may be months stale, we show the specification and send you to the retailer, who knows today's price. Budget is your ceiling, not our guess at anyone's price.

When the answer is "wait"

Memory and graphics cards are unusually expensive in 2026: AI datacentre demand redirected the capacity that used to make consumer parts. Where that affects the part you would need, and you can run the model some other way meanwhile, we say so and give the date relief is expected, with our sources. A rumour is labelled a rumour, and we tell you not to defer a purchase on one.