Hardware guide
You do not need a flagship card to run local models well. You need memory — and the used market is full of cards whose memory outlived their speed. This page names the ones worth buying, the ones worth avoiding, and what to check before money changes hands. It names no prices: street prices move weekly, and this site only publishes numbers it can measure on a schedule.
Buy memory, not speed. Twelve gigabytes is the honest floor, sixteen is the comfortable middle, twenty-four is where the used market hands you a previous-generation flagship for the price of a new mid-range card. Everything below that runs the small models and nothing comfortable above them.
◇ Estimated— Estimated: computed from our curated model and hardware catalog — not a live reading.Counts are the tracked models rated Excellent or Good on each card at 8K context, computed by the same fit engine as every hardware page.
12 GB — the entry point that works
The used-market king of local AI.
Twelve gigabytes was an oddly generous choice for a mid-range card, and it is the whole reason this card matters: the popular 4-bit quantizations of 7–14B models fit with room for a real context. It is slow by modern standards and that is fine — chat is interactive at this size. If you want the cheapest honest start, this is it.
14 of 45 tracked models run comfortably · largest: Gemma 4 12B (~6.6 GB Q4_K_M)
16 GB — one size up, quietly
The step that admits the 20–30B class.
Sixteen gigabytes opens the models that feel noticeably smarter than the 7–14B tier — the mid-size mixture-of-experts releases and the 24B dense models at 4-bit. Bandwidth is modest, so generation is unhurried, but the models it runs are the ones many people settle on for daily work.
16 of 45 tracked models run comfortably · largest: gpt-oss 20B (~10.8 GB Q4_K_M)
24 GB — the value jump
A previous-generation flagship with today's flagship memory.
Twenty-four gigabytes is the same memory as the current top consumer card, on a chip two generations old — which is exactly what local AI wants, because memory decides what fits and bandwidth only decides how fast. The 30B class becomes comfortable, and the used market is where this card lives now. Expect a big, hot, power-hungry card; expect it to be worth it.
22 of 45 tracked models run comfortably · largest: GLM-4.7-Flash (~17.1 GB Q4_K_M)
Unified memory — the other route
Not a GPU at all, and a serious contender.
A Mac with 64 GB of unified memory plans with roughly 45 GB for a model — territory no consumer graphics card reaches. Generation is slower than a big discrete GPU and the machine is silent, small and efficient. If you already own one, run the finder before buying anything.
29 of 45 tracked models run comfortably · largest: Qwen AgentWorld 35B-A3B (~20.6 GB Q4_K_M)
Eight gigabytes runs the small models (up to ~8B at 4-bit) and nothing comfortable above them. A fast 8 GB card loses to a slow 12 GB card for this workload every time, because memory decides what fits at all. If a listing shouts about clock speeds and whispers about memory, walk on.
A "4090" in a laptop carries 16 GB, not 24 — the desktop name on a different chip. Judge a laptop by the memory figure in its spec sheet, never by the model number, and expect the finder's verdicts to move down a tier.
The 100B+ releases at the top of the momentum charts need data-center memory or a rented GPU. Buying consumer hardware to chase them ends in offloading and disappointment. Pick the card for the models you will actually run daily; rent an hour of a big GPU when you are curious.
Two cheap cards do not simply add up: the runtime has to split the model, the slower link between them costs speed, and the power and case requirements double. One card with more memory beats two with less, until you genuinely outgrow the largest single card.
Run the finder against the card you already have — many people discover their current GPU runs the models they wanted. Then pick a runtime and try a model that fills the memory. A budget is easiest to keep when the first purchase is a download.
Check what fits your card →Pick a runtime →VRAM, explained →