Choosing an external SSD for storing AI models on a Mac
Affiliate disclosure: Links labeled “(paid link)” may earn Tried AI Tools a commission if you sign up or buy through them, at no extra cost to you. Commissions do not decide what we list or recommend. As an Amazon Associate, Tried AI Tools earns from qualifying purchases. How this site is funded
Short answer
Buy a Thunderbolt or USB4 SSD, or a USB4 enclosure with an NVMe drive, for about 3,000 MB/s on any recent Mac, or a Thunderbolt 5 SSD for even faster loads on Macs with Thunderbolt 5 ports. Avoid paying extra for USB 3.2 Gen 2x2 (20 Gbps) drives, because Macs run them at 10 Gbps, about 1,000 MB/s.
Local AI models are big. A 35B model is about 24 GB and a 120B model about 65 GB, and most people end up keeping several. An external SSD is far cheaper than upgrading a Mac's internal storage at purchase, but not every fast-looking drive is fast on a Mac.
Why drive speed matters (a little)
The drive only matters when a model loads. Load time is roughly model size ÷ drive speed:
| Model size | ~1,000 MB/s (USB 10 Gbps) | ~3,000 MB/s (Thunderbolt 3/4, USB4) | ~6,000 MB/s (Thunderbolt 5) |
|---|---|---|---|
| 8 GB | ~8 s | ~3 s | ~1.5 s |
| 24 GB | ~24 s | ~8 s | ~4 s |
| 65 GB | ~65 s | ~22 s | ~11 s |
If you switch models often, a faster drive saves real time. If you load one model in the morning and keep it all day, a 10 Gbps drive is fine.
The USB 3.2 Gen 2x2 trap
Many portable SSDs advertise 2,000 MB/s using USB 3.2 Gen 2x2 (20 Gbps). Macs do not support that mode, so these drives fall back to 10 Gbps and run at about 1,000 MB/s, the same as cheaper drives. Check the connection type, not just the headline speed.
The three sensible choices
| Type | Real speed on a Mac | Best for |
|---|---|---|
| USB 3.2 Gen 2 (10 Gbps) portable SSD | ~1,000 MB/s | Budget, occasional model switching |
| Thunderbolt 3/4 or USB4 SSD, or a USB4 enclosure with an NVMe drive | ~2,800 to 3,800 MB/s | Most people; best value per gigabyte at speed |
| Thunderbolt 5 SSD | ~5,000 to 6,000+ MB/s | Macs with Thunderbolt 5 ports (such as the 2026 Mac Studio), frequent switching between large models |
A USB4 enclosure plus an NVMe SSD you choose yourself is often the best deal: you get Thunderbolt-class speed and can pick capacity freely.
What else to look for
- Capacity: 2 TB at minimum for a model library, 4 TB if you also keep image models.
- Cooling: sustained reads of tens of gigabytes can make small drives throttle. Metal enclosures with heatsinks hold their speed better.
- Cable: use the short cable that came with the drive; a long or cheap cable can drop the connection to a slower mode.
- Format: format the drive as APFS in Disk Utility for the best performance on a Mac.
Moving your models to the drive
- Ollama: quit Ollama, move the
~/.ollama/modelsfolder to the SSD, then set the location withlaunchctl setenv OLLAMA_MODELS /Volumes/YourDrive/ollama-modelsand restart Ollama. - LM Studio: open the My Models tab and change the models directory to a folder on the SSD.
For how much model you can actually run, see how much memory you need.
Shop on Amazon
Before you buy, check that the drive or enclosure says Thunderbolt or USB4. A USB 3.2 Gen 2x2 drive runs at about 1,000 MB/s on a Mac.
- Thunderbolt 5 SSDs (paid link)
- Thunderbolt and USB4 SSDs (paid link)
- USB4 NVMe enclosures (paid link)
- NVMe SSDs to put in an enclosure (paid link)
These links open Amazon search results. We do not vouch for any one seller or listing, so check the specs, seller and return policy before you buy.
Frequently asked questions
Can I run models directly from an external SSD?
Yes. The drive's speed only affects how long a model takes to load into memory. Once loaded, the model runs from unified memory at full speed. Point Ollama at the drive with the OLLAMA_MODELS environment variable, or change the models folder in LM Studio's settings.
How big a drive do I need?
Models range from a few gigabytes to over 80 GB each. A 2 TB drive holds a comfortable library of 20 to 40 models; 4 TB suits people who keep several large models and image models.