Count the models you keep locally, add an allowance for the variants and datasets that accumulate around them, and the calculator returns an approximate library size plus the capacity tier that covers it. Every model figure is an approximate published size at Q4-class quantization and varies by quantization.
Add the models you keep locally and this will size the drive. All figures are approximate.
Almost everyone sizes this by taking the largest model they intend to run and adding a bit. That method fails within a month, and it fails for reasons that are structural rather than careless. The largest model is never the problem; the accumulation around it is.
The first multiplier is quantization. A model family you think of as one item is, in practice, several files. You download a Q4 variant because that is the common recommendation, then a Q5 because you want to see whether the difference matters to you, and possibly a Q8 or the original weights before you understood the tradeoff. Nobody deletes the losers immediately. A 13B family with an approximate ~8GB Q4 footprint routinely occupies three or four times that once you have finished comparing.
The second multiplier is families. Local AI is a field where new releases arrive constantly and trying them is the point. Each one you try is a permanent resident until you deliberately clean up, and deliberate cleanup is a task everyone postpones. The library grows monotonically because deleting requires a decision and downloading does not.
The third category is everything that is not a model at all, and it is the one no sizing guide counts. Fine-tuning datasets. LoRA adapters, which are individually small and collectively not. Generated output, which for image work becomes the largest single directory on the machine within months. Container images for each runtime you evaluated, where the layers persist long after you stopped using the tool. Your own training checkpoints, which are both large and the only files here you cannot re-download.
The headroom slider above defaults to fifty percent for this reason. It is not padding for its own sake; it is an allowance for the three multipliers, and in our view it is a floor rather than a generous cushion. If the calculator says you need a terabyte, the honest planning number is closer to a terabyte and a half.
The calculator gives you one total, but that total is not all one purchase, and treating it as one is how people end up buying far more flash than they need. The number that matters for your NVMe drive is the working set: the models you will actually load in a given week or month. For most people that is a small fraction of the library.
A reasonable way to divide it is by whether you would notice the file being gone. Models you load regularly are working set. Models you keep because you might come back to them, older quantizations you have superseded, datasets you have finished with, and generated output you are not currently using are all archive. The first group sizes your NVMe purchase; the second sizes your disk purchase.
The split is worth doing explicitly because the two tiers differ by roughly an order of magnitude in cost per terabyte at today's tracked prices. Getting the division wrong in the expensive direction is the single most common way to overspend in this category. Getting it wrong in the cheap direction costs you nothing but a file copy when you want something back.
If your working set genuinely is most of your library — which happens for people who curate aggressively and delete what they stop using — then you do not need an archive tier and should simply buy the NVMe capacity. The two-tier architecture is a cost optimization, not a requirement, and if it does not save you money in your usage pattern then skip it.
Running out of space on a working drive is normal and does not mean the original purchase was wrong. There are three sensible responses and one bad one, and knowing which is which in advance saves an expensive panic purchase.
The first response is to move the archive off. If you have been keeping everything on one drive, the cheapest recovery is to buy disk capacity and relocate the cold portion of the library. This is almost always the correct first move because it is the cheapest per terabyte and because it addresses the actual growth, which is in the archive rather than the working set.
The second is to add a second NVMe drive rather than replace the first. Desktop boards commonly have more than one M.2 slot, and two drives cost the same per terabyte as one of double the size, sometimes less. Replacing a working drive means migrating everything and retiring hardware that is fine; adding one means plugging it in.
The third, once the library is genuinely large or you want it available to more than one machine, is a network share. That is a real cost step up and it is covered properly on the model library NAS page, including what a live build actually costs today. Our NAS full guidance covers the adjacent case of an existing array running out of room.
The bad response is buying the largest flash drive available in a hurry. Capacity bought under time pressure is capacity bought at whatever price the market happens to be asking, and in a market documented as supply-constrained that is a bad moment to be a forced buyer.
One useful way to sanity-check a storage decision is to price the models themselves rather than the drive. The arithmetic is trivial and the result is often clarifying: take the live cost per terabyte from the table below and multiply by the approximate footprint of the model.
At the cheapest in-stock NVMe price we are tracking right now, keeping a ~40GB 70B-class model on flash costs a few dollars of drive space. Keeping the same model on archive disk costs a fraction of that. Neither number is large in isolation, which is exactly the point — individual models are cheap to store, and that is why libraries grow without anyone noticing until the drive is full.
The number that is not small is the total. A few dozen checkpoints and quantization variants at those per-item costs adds up to a real drive purchase, and the difference between doing that on flash and doing it on disk is the difference between a small expense and a large one. This is arithmetic from live prices rather than a performance measurement or a projection: the inputs are the current cheapest in-stock price per terabyte and the published approximate model sizes.
It is also the argument for not over-buying flash up front. If archive capacity costs a small fraction per terabyte of what flash costs, then the correct response to an uncertain future library size is a modest working drive now plus cheap capacity when you need it, rather than a large flash purchase against a library you have not built yet.
The output above is a capacity tier, not a product, and the distinction is deliberate. Tiers are how drives are actually sold — one, two, four and eight terabytes — and the cost per terabyte at each tier moves independently with the market. Recommending a tier and then showing you the cheapest in-stock drive that clears it separates the sizing question, which is stable, from the pricing question, which is not.
Treat the total as approximate in both directions. It is built from published approximate model footprints, so it inherits their imprecision, and it cannot know how your particular usage will grow. What it is good for is deciding between adjacent tiers: if the number lands close to a boundary, the headroom slider will tell you quickly whether you are a two-terabyte buyer or a four-terabyte one, and that is the decision that actually costs money.
If the recommended tier looks larger than you expected, check the headroom setting before assuming the calculator is wrong. Fifty percent on a large library is a large absolute number, and it is doing real work — but it is a planning allowance rather than a measured requirement, and lowering it is a legitimate choice if you delete aggressively.
If it looks smaller than you expected, that is worth sitting with. The most common storage mistake in local AI is buying flash capacity against a library you have not built yet, on the assumption that AI work requires enormous storage. For the working set, it usually does not. The archive tier is where the terabytes go, and archive terabytes cost a fraction of flash ones.
Live prices · updated every 4-5 hours · last checked 19 min ago · ranked by price per terabyte
| Drive | Capacity | PCIe Gen | Read MB/s | $/TB | Price | |
|---|---|---|---|---|---|---|
| Seagate Nytro 5060 U.2 7.68TB | 7.68TB | Gen4 | — | $23.73 | $182 | Buy |
| Samsung PM9A3 3.84TB U.2 NVMe | 3.84TB | Gen4 | — | $25.45 | $98 | Buy |
| WD Ultrastar DC SN655 7.68TB U.2 | 7.68TB | Gen4 | — | $34.45 | $265 | Buy |
| WD WD_BLACK SN850X 1TB NVMe SSD | 1TB | Gen4 | 7,300 | $100.00 | $200 | Buy |
| Fikwot FX660 4TB M.2 SSD | 4TB | Gen4 | 5,000 | $109.25 | $437 | Buy |
| Solidigm D5-P5336 7.68TB NVMe U.2 | 7.68TB | Gen4 | — | $117.19 | $900 | Buy |
| Silicon Power 4TB US75 Nvme PCIe Gen4 M.2 2280 SSD R/W Up to 7,000/6,500 MB/s with | 4TB | Gen4 | 6,500 | $119.99 | $480 | Buy |
| Ediloca 4TB PS5 SSD with Heatsink PCIe Gen4 NVMe M.2 Gaming SSD, 7400MB/s | 4TB | Gen4 | 7,400 | $120.00 | $480 | Buy |
They are widely published approximate figures at Q4-class quantization and every one of them is labeled with a tilde for that reason. Real files vary by quantization level, by how a particular release was packaged, and between GGUF builds of the same weights. Use them for planning a drive purchase, not for provisioning to the gigabyte.
Because libraries grow through quantization variants, new model families, datasets, container layers and generated output — none of which appear in a naive model-count estimate. Fifty percent is our floor rather than a cushion. You can lower it with the slider, but we would not.
Buy the cheapest in-stock drive at or above it. Capacity tiers are coarse — 1TB, 2TB, 4TB, 8TB — and the cost per terabyte at each tier does not rise smoothly, so the next size up is sometimes cheaper per terabyte than the one you calculated. The live table shows which is which today.
No. The calculator sizes your model library only. If the drive you are buying is also your boot drive, add whatever your system and applications occupy — for most desktop installs that is a meaningful fraction of a terabyte before any models exist.
That is what headroom is for. The most reliable predictor of next year's library size is that it will be larger than this year's by more than you expect, because trying new releases is the entire appeal of running models locally.
For the archive tier, yes — it is cold capacity and an external enclosure is a legitimate way to buy it. For the working set it is a question of convenience and connection rather than one we would answer with performance claims. Our external drives coverage tracks live prices on that side.