NVMe SSDs for Ollama models and GGUF files

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

Local language models through Ollama take gigabytes on disk. Each GGUF variant (the compressed weight format) keeps its own copy of part of the model. Whisper (open-source speech-to-text) and ComfyUI with FLUX (a family of image-generation models) or SD (Stable Diffusion, open-source image generation) add more heavy folders.

The short version: a text-only profile often fits on a 512 GB NVMe. A voice, images, and backups profile should aim at 1–2 TB. This guide follows four steps: size, buy, move, and maintain. It does not promise a speed figure, and it does not invent a price.

Guide contents

Who this is for, and where to start

Reader Path Rough time What you need first
Beginner (a first ollama pull) The tables below, plus steps 1–3 of the move path ~20 min of reading A first local AI in 15 minutes, or Ollama already installed
Intermediate (a family, ComfyUI) Collapsible steps 4–6, plus the quarterly checklist 1–2 hours to set up Admin access on Windows or Linux for OLLAMA_MODELS

Size the drive

What is on the disk Indicative GB Example
3 chat GGUF models 15–25 Mistral 7B, Llama 8B, a small French-language model
Whisper (small or medium) 1–3 Offline transcription
FLUX / SD plus checkpoints 20–60 Images without a cloud
Docker images 10–30 Docker and Ollama
OS plus spare room 80–120 Windows or Linux, plus updates

Text only: 512 GB can be enough. Text, voice, and images: aim for 1–2 TB. Use VRAM for graphics memory. A fast SSD does not fill in for too little system RAM, 32 vs 64 GB. Swap wears the NVMe.

Household profiles

Profile Suggested capacity Comment
Trying it out 512 GB NVMe Two models plus the OS. See a first local AI in 15 minutes
Family chat 1 TB Open WebUI plus 3–4 GGUF files
Creative work plus voice 2 TB ComfyUI, Whisper, and backups

A few words. FLUX and SD are heavy checkpoints for local ComfyUI. Docker images are the layers for Whisper or n8n. docker system prune leaves Open WebUI volumes in place.

Common storage mistakes

What you see Likely cause What to try
Drive C: fills up after two pull commands Models landed in %USERPROFILE%\.ollama Point OLLAMA_MODELS at a second SSD. Windows guide
Inference is slow from the NAS The PC is reading over the network Keep active GGUF files local. Use the NAS for archives (Synology)
“No space left” on Linux The / partition is too small A dedicated mount such as /srv/ai-models (step 3 of the move path)
Docker and native Ollama double the gigabytes Two copies of the same tags Pick native or Docker, and run both only when you have a reason (Docker and Ollama)
ComfyUI can no longer find a checkpoint A folder was moved by hand Write the paths in a README.txt inside AI-Models

A spinning hard drive stretches load times. Skip it for everyday inference. A Gen4 NVMe shows its speed mainly on the first ollama run after a reboot. Token speed still comes from the CPU or GPU (LM Studio GPU offload, VRAM).

Buy the right NVMe

Criterion Why
Capacity Follow the profile table. The Gen4 label is secondary
PCIe Gen3 vs Gen4 Gen4 means faster first loads. Once the model is in RAM, the gap is small
Endurance (TBW) Writes add up if you try a lot of models
DRAM cache More comfortable when you copy large files
Gen3 vs Gen4, and a second M.2 slot (collapsible, optional)

Gen3 is enough for inference once the model is in RAM. Gen4 mainly speeds the first load and copies of 20–60 GB. One M.2 slot means you should aim at 1–2 TB for everyday use, and keep USB for backups. Upgrade order: RAM, then SSD capacity (the hardware guide).

How to buy (without overbuying)

  1. Read the household profile that matches you (512 GB, 1 TB, or 2 TB).
  2. Check the M.2 slot and case compatibility (the manufacturer page, or the mini PC comparison).
  3. Open one search per capacity below. Choose 1 TB if you combine chat and Whisper, and 2 TB if ComfyUI or FLUX is in the picture. Check today’s price and stock (October 2026). LocAIhome does not publish a fixed price.
  4. Cross-check the hardware selector if you are choosing between a new mini PC and a drive upgrade on the PC you have. The tool filters budget, noise, and uses. Buying is optional. See how affiliation works.

Indicative search: 1 TB NVMe (chat plus Whisper)

amazon.ca affiliate link: 1 TB PCIe NVMe SSD: an open search. A commission is possible at no extra cost to you. How affiliation works. Check M.2 compatibility (2280, PCIe) before you order.

Indicative search: 2 TB NVMe (creative work, voice, and backups)

amazon.ca affiliate link: 2 TB PCIe NVMe SSD: the same affiliate frame as above. Affiliation. A fit when you combine ComfyUI or FLUX with several GGUF files.

Move the files and organize storage

Before any ollama pull, pick the Windows or Linux column. The NAS is for archives, not for everyday chat.

OS Folder for models you run Archives / backup
Windows D:\AI-Models\ollama, or another drive letter off the system drive C: An encrypted external drive (BitLocker on a dedicated volume)
Linux /srv/ai-models/ollama on a local NVMe rsync to a Synology NAS
NAS Everyday inference from the network is a poor fit Cold storage for old GGUF files. Keep 2–3 active models on the PC

If you are not the person who installs software. The paths D:\ and /srv/ are examples. You do not need to understand partitioning to size the drive. If someone else installs Ollama for you, ask only “where are the models stored?” and write the answer down for the private family assistant (several people on the same LAN).

One path, steps 1–6

  1. Take inventory. Run ollama list, and add Whisper weights and ComfyUI checkpoints if you use them. Add a 25% margin. Skip duplicate Q4 and Q5 copies of the same model: choose a first model.
  2. Pick the capacity. Under 400 GB of files you actually need → a 512 GB NVMe. From 400 to 900 GB → 1 TB. Voice plus images → 2 TB (see the household profiles).
  3. Create the folder. Windows: D:\AI-Models\ollama. Linux: /srv/ai-models/ollama. Creating the folder is enough. Advanced partitioning is optional.
Steps 4 to 6: point Ollama at the folder, archive, and check (collapsible)
  1. Point Ollama at the folder. Set the OLLAMA_MODELS variable (Ollama’s docs). Windows, so it survives a reboot: install Ollama on Windows. Linux: set the variable in the systemd service, then sudo systemctl restart ollama.
  2. Separate what you run from what you archive. Inference stays on the local NVMe. Old GGUF files go to an encrypted NAS (privacy). Free more than 30 GB on C: or / before an ollama pull of a 7B Q4. Use ollama rm on the experiments.
  3. Check. ollama list looks right. The path is written down for the family. Names include the version and the quantization (for example 7B-Q4). An optional disk migration is a clone (Macrium, Clonezilla). Keep the old drive for a week. Advanced Linux (a second NVMe, fstab) needs a backup first. Otherwise stay on Windows, or follow install Ollama on Windows, Linux, and Mac.

What you should see. ollama list shows your models. New pull commands write into the folder you chose, and they leave a full C: alone. The family knows where the GGUF files are for Open WebUI or the private assistant.

Keep it tidy

Backup and encryption

OS Usual location What to do
Windows %USERPROFILE%\.ollama or OLLAMA_MODELS Copy to an external drive, or to an encrypted partition off C: (BitLocker on that backup volume). Leave the system disk as it is
Linux ~/.ollama or OLLAMA_MODELS rsync to a Synology NAS
macOS ~/.ollama Time Machine on the Mac. See Mac mini M4 for local AI for how an Apple Silicon disk is laid out

TBW (terabytes written, the endurance the maker warrants): read the manufacturer page. For 3–4 models you keep, capacity (1–2 TB) matters more than an extreme TBW margin.

Quarterly checklist (maintenance)

Once a quarter, in about 15 minutes:

  1. ollama list: note the models you actually use, and ollama rm the duplicates.
  2. The ComfyUI models/ folder, if you generate images (images without a cloud).
  3. The Whisper cache (~/.cache/whisper). See Whisper offline.
  4. Docker: docker system prune frees images. It leaves Open WebUI volumes in place. Detail is in Docker and Ollama.
  5. Line the disk up with the bill: the electricity budget if the machine runs 24 hours a day (mini PC noise).
Optional M0–M24 plan (collapsible)

Quarterly markers: M0, 512 GB or 1 TB from the household profile. M6, ollama list, and remove the experiments. M12, Whisper or ComfyUI, and move to 1–2 TB if you need it. M24, archives on the NAS, with 2–3 active GGUF files on the NVMe.

Go further

The drive is often the second upgrade after RAM or a GPU, especially if Ollama on Windows is already running on the PC you have. The hardware selector lines up storage, VRAM, and the mini PC comparison. The Amazon boxes above open an affiliate search for a capacity. You pick the drive from the specs on this page.

Common questions

Is 512 GB enough to start?

Yes for two or three text models. Plan on 1 TB if you add Whisper, ComfyUI or FLUX, or backups.

Is NVMe Gen4 required?

Gen3 is often enough for inference. Gen4 mainly shortens the first load.

Can I store the models on a NAS?

Yes as an archive. Inference over the network is slow. Keep the models you actually run on a local NVMe drive.

How much space for three chat models plus Whisper?

About 15–25 GB for three chat GGUF files, 1–3 GB for Whisper, plus room for the OS. See the “Size the drive” table at the top of the guide.

Sources

Narrow it down with the hardware selector · French version of this section