Every machine that runs local models, priced

All 56 configurations this site lists, in one table: what each costs, how much of its memory the GPU can use, how many of the 50 current models it holds at 32k of context, the strongest of those, and how long that pair takes to pay for itself rather than renting the same model. 54 of them carry a published price. The Mac Studio M5 Ultra, 512GB and the Framework Desktop, 192GB do not, so their rows open the calculator for you to put in what you would pay. Families are in order of what the cheapest of them costs. For the quickest pay-back at a given amount of use, see best buys by usage; for two machines side by side, every head-to-head; for what each model needs before you pick a box, how much memory you need.

Jump to: Does a dearer machine run a better model? · How long each machine takes to pay for itself · Every machine here, side by side · The assumptions behind the table

Does a dearer machine run a better model?

The short version: more money buys memory, and memory buys a stronger model in only five steps. Of the 56 machines here, 39 top out at the same model, Qwen3.8 27B: everything from the Mac mini M4, 32GB at $999 to the RTX PRO 6000 Blackwell, 96GB at $18,000, card only. Five hold something stronger, and the cheapest of those is the Mac Studio M3 Ultra, 256GB at $7,099, while two hold all 50 current models. Between those steps the money buys speed, spare memory and a longer window rather than a better model. The Mac mini M4, 32GB and the Mac Studio M3 Ultra, 256GB are discontinued, so their prices are the ones they launched at.

How long each machine takes to pay for itself

Pay-back runs the other way, because the strongest model a machine holds is never the fastest thing it can run. At 500k tokens a day on that model, the quickest figure in the table is 26 years, on the GeForce RTX 3060, 12GB, and the slowest is 2,379 years. Nothing here pays for itself inside a decade at that usage. What changes the answer is using the machine much harder. Swapping the model changes it on every one of the 54 machines that pay back at all: each of them holds something that pays back sooner than the model in the table. Best buys by usage ranks every pair. The GeForce RTX 3060, 12GB is discontinued, so its price is the one it launched at.

Every machine here, side by side

The assumptions behind the table

Usable memory is what the GPU can address, which is less than the memory fitted: on a Mac it follows the macOS wired limit, and on a graphics card it is the VRAM less the gigabyte llama.cpp leaves free. The count is of the 50 current open models at 32k; ask for a longer window and the cache grows, so fewer fit, and each machine's own page gives the length it takes every model to. Pay-back is that machine running the strongest model it holds, at 500k tokens a day, 15:1 input to output, $0.17 per kWh, and today's API prices held flat; best buys ranks every machine-and-model pair by how soon it pays back instead. Graphics cards are priced as the card alone, so add the PC around one before comparing it with a complete computer. All seven cards here are ranked by what each one holds. A discontinued machine is priced at what it launched at, which is not a price you can pay today. Each price opens the calculator on that machine running the model beside it.