How Much SSD Do You Need for a Local Model Library?

Bottom Line

• Rule of thumb: size to your VRAM, then multiply by three. The multiplier is quantization variants, not model count.
• 1TB is the floor. ~700–800GB free after everything else. Fine for three or four 24GB-tier models. It will fill.
• 2TB is the right default for a single-GPU box. 4TB once you are at 48GB or more of VRAM.
• A faster SSD does not make your tokens faster. Weights load once into VRAM. After that the disk is idle.
• It does change load time — a few seconds on NVMe, roughly a minute on SATA, minutes on a hard drive, for a 40GB model.
• The thing that actually makes reloads instant is system RAM, not drive class, because the OS caches the file.
• Prices are unstable. NVMe averaged +113.7% between November 2025 and March 2026. Shop on dollars per terabyte.

Real Model Footprints

These are figures we have published elsewhere on this site and keep consistent, not estimates.

Model	Build	On disk
gpt-oss 20B	Q4_K_M	~12–13 GB
Qwen 3.6 27B	Q4_K_M	~17–18 GB
Laguna XS 2.1	Q4_K_M (default Ollama tag)	20.27 GB
gpt-oss 20B	Q8_0	~21 GB
Qwen 3.6 27B	Q6_K	~22–24 GB
Dense 32B	Q8_0	~35 GB
Laguna S 2.1	UD-Q2_K_XL	39.7 GB
Dense 70B	Q4_K_M	~40 GB
Laguna S 2.1	UD-IQ4_XS	57.6 GB
Laguna S 2.1	UD-Q4_K_M	73.1 GB
Kimi K2.7-Code	~2-bit	~340 GB
GLM-5.2	~2-bit	~256 GB
Kimi K3	native MXFP4	~1.4 TB

That last row is not a joke entry. It is a real 2026 release you cannot store on a consumer machine, let alone run — and it is the direction frontier open weights are moving. See open weights are not local any more.

The Hoarding Multiplier

Here is where every sizing guide goes wrong. It counts models. You do not download models — you download candidates.

A realistic six-month library on a 24GB card:

What you downloaded	Why it is still there	Size
Laguna XS 2.1 Q4_K_M	Your daily driver	20.3 GB
Laguna XS 2.1 Q6_K	You wanted to compare quality	~26 GB
Qwen 3.6 27B Q4_K_M	The alternative you tested	~17.5 GB
gpt-oss 20B Q8_0	Better at one specific task	~21 GB
gpt-oss 20B Q4_K_M	Downloaded first, never deleted	~12.5 GB
An embedding model	RAG experiment	~1 GB
A previous release	You will “get to it”	~20 GB
Total		~118 GB

That is a modest library, at the smallest VRAM tier that matters, and it is already 118GB. Run the same pattern on a 48GB card, where the interesting builds are 40–73GB each, and you are past 400GB before you have made any unusual decisions.

The multiplier is roughly 3x your VRAM per model you seriously evaluate. Q4, Q6 and Q8 of the same model is three copies. That is the number to plan around, and it is why “1TB NVMe” appears in every parts list and satisfies almost nobody. If quantization labels are new to you, start with quantization in plain English.

Sizing Table

Assume the operating system, applications and ordinary files take 200–300GB.

Drive	Free for models	Comfortable for	Verdict
1TB	~700–800 GB	3–5 models at 12–24GB	The floor. Works, needs discipline.
2TB	~1.7–1.8 TB	A full 24GB-tier library, or 4–5 models at 40GB	The right default.
4TB	~3.7–3.8 TB	48GB+ tier, several 70B-class builds, fine-tuning outputs	Buy this if VRAM ≥ 48GB.
8TB+	—	Multi-model serving, dataset work, 200GB+ builds	A working-storage decision, not a model-library one

Fine-tuning changes the answer entirely. Checkpoints, optimizer states and dataset copies dwarf the base model, and they arrive in multiples. If training is in your plan, treat the table above as the floor and add a separate scratch drive.

Does Drive Speed Matter? The Arithmetic

Load time is size divided by sequential read speed. These are ceiling figures at rated speed — real loads are slower because of decompression, metadata and the loader itself. Use them as ratios, not stopwatch predictions.

Drive class	Rated sequential read	20 GB model	40 GB model	73 GB model
Hard drive	~150 MB/s	~2.2 min	~4.4 min	~8.1 min
SATA SSD	~550 MB/s	~36 s	~73 s	~2.2 min
NVMe Gen4	~7,000 MB/s	~3 s	~6 s	~10 s
NVMe Gen5	~14,000 MB/s	~1.4 s	~3 s	~5 s

The real gap is HDD to SSD, not Gen4 to Gen5. Going from a hard drive to any NVMe drive turns minutes into seconds. Going from Gen4 to Gen5 turns six seconds into three, on an operation you perform once a session. Gen5 drives cost substantially more per terabyte and run hot enough to want a heatsink. For a model library, that money buys more capacity on Gen4 than it buys speed on Gen5.

The Part That Actually Decides Your Load Times

Your operating system caches file reads in free system RAM. Load a 20GB model once on a machine with plenty of spare RAM, and the second load comes from memory rather than disk — effectively instant, regardless of what drive it sits on.

So the honest ranking of what governs perceived load speed:

1. Do you have enough free system RAM to cache the model? If yes, the drive stops mattering after the first load.
2. How often do you swap models? If you run one model all day, you pay the disk cost once at boot. If you use a hot-swapping setup that unloads idle models — which is exactly what a multi-model agent stack does — you hit the disk constantly, and now the drive class is real.
3. What class is the drive? Third, not first.

This inverts the usual advice. “Buy a fast SSD so models load quickly” is aimed at case 2 and quietly wrong for case 1. If you run a single model and have RAM to spare, spend the money on capacity. If you swap between five models across the day, spend it on the drive.

And to answer the question underneath all of this plainly: once weights are in VRAM, the disk plays no part in generating tokens. A faster SSD will not improve your tokens per second or your first-token latency. Those are governed by memory bandwidth and by how much context you can hold.

Where the Files Live, and How to Move Them

Boot drives fill first because everything defaults there.

Tool	Default location	Override
Ollama	~/.ollama/models	OLLAMA_MODELS
Hugging Face	~/.cache/huggingface/hub	HF_HOME
llama.cpp	Wherever you saved the GGUF	Path argument

Pointing these at a second large drive is much cheaper than replacing a boot drive. One gotcha worth stating: exporting a variable in your shell does not reach a service. If Ollama runs under systemd, set it in the unit or a drop-in file, or the daemon will keep writing to the old path and you will conclude the setting does not work.

What to Buy

Prices are genuinely unstable and we are going to be specific about what we do and do not know.

The NAND shortage has been severe. GamersNexus tracked four common 2TB Gen4 NVMe drives from November 2025 to March 2026 and measured the average going from $168.75 to $357.50 — up 113.7% — with individual models between +84% and +153%. Our own August 2026 reference puts budget 1TB Gen4 drives at $90–165.

Those two figures do not scale into each other, and we are not going to pretend they do. They cover different tiers at different dates. The useful conclusion is that the spread between budget and premium drives at the same capacity is now wider than the gap between capacities — so shop on dollars per terabyte and check current listings rather than trusting any single number, including ours.

We could not verify a 4TB street price from a source we trust, so we are not printing one. The aggregator trackers we found disagreed with themselves within the same page. Check listings.

SanDisk 1TB NVMe SSD PCIe 4.0	The floor. Fine as a dedicated model drive alongside an existing boot drive.
SanDisk 2TB NVMe SSD PCIe 4.0	The default recommendation for a single-GPU local AI box.
SanDisk 4TB NVMe SSD PCIe 4.0	For 48GB+ VRAM machines, where single builds run 40–73GB.

One caveat we have not seen anywhere else. Building a model library is an unusually write-heavy consumer workload. Filling a 4TB drive twice writes 8TB. Most desktop use never approaches a drive’s rated endurance; downloading and deleting large quantizations for a year is one of the few things that can move the needle. Check the TBW rating on whatever you buy, and prefer a drive that publishes one.

External storage is the right archive tier. USB4 and Thunderbolt run at 40 Gbps, USB 3.2 Gen 2 at 10 Gbps, and real throughput sits below both ceilings — so a USB4 NVMe enclosure is close enough to internal Gen4 for load times, while a 10 Gbps enclosure is several times slower. Keep the two or three models you actually run on the internal drive and archive the rest.

SABRENT USB4 NVMe enclosure (tool-free, 3,900 MB/s) — tool-free, M.2 2280, 40 Gbps. This is the archive tier in one part. Pair it with a drive that publishes a TBW rating, per the caveat above.

The Decision

Your situation	Answer
12–16GB VRAM, one or two models	1TB, with discipline
24GB VRAM, normal experimentation	2TB
48GB+ VRAM	4TB
You hot-swap between many models	Spend on the drive — you are disk-bound
One model, plenty of system RAM	Spend on capacity — the OS cache hides the drive
Fine-tuning	4TB minimum, plus separate scratch space
Drive already full	Set OLLAMA_MODELS / HF_HOME, add a second drive
Considering Gen5 for speed	Buy more Gen4 capacity instead

See Also

• How to Move Ollama Models to Another Drive — the step-by-step for OLLAMA_MODELS on each operating system
• Can You Run Kimi K3 Locally? — the 1.4TB model that makes this page necessary
• Open Weights Are Not Local Any More — why footprints keep growing
• Quantization in Plain English — what Q4, Q6 and Q8 mean and why you end up with all three
• Is 48GB of VRAM Enough in 2026? — the tier where 4TB becomes the right answer
• How Much VRAM for 128K Context? — the memory that actually governs speed
• Motherboard and CPU for a Multi-GPU LLM Rig — M.2 slot counts and lane sharing

Sources

• Model footprints are our own published figures, kept consistent across the site — Laguna XS 2.1 at 20.27GB and Laguna S 2.1 at 39.7 / 57.6 / 73.1GB from can you run Kimi K3 locally; 20B–70B figures from our VRAM tier pages
• Load-time figures are our own arithmetic — capacity divided by rated sequential read — and are ceilings, not measurements
• SSD pricing: GamersNexus, “SSDs: WTF?”, published 2026-04-02 — four 2TB Gen4 NVMe drives averaging $168.75 in November 2025 and $357.50 in March 2026, +113.7%. Budget 1TB Gen4 at $90–165 as of August 2026 from our own hardware price reference (Tom’s Hardware SSD tracker, Kingston and WD listings)
• No 4TB price is quoted because we could not verify one from a source we trust
• USB4/Thunderbolt 40 Gbps and USB 3.2 Gen 2 10 Gbps are interface specifications, not measured throughput