Every “AI on a USB drive” video I’ve watched gives the same advice: buy a fast drive, at least USB 3. It makes sense. If the model lives on the drive, a faster drive should mean a faster AI. I wanted to see how much the drive really changes, so I put the same app and the same model on a USB SSD and on my Mac’s own disk and timed everything.
📺 Watch the video: AI on a USB Drive: I Tested If It Actually Slows You Down
What’s on the drive
The app is Portable Local Studio by Tech Jarves, the one from his USB drive video. It bundles four open-source engines behind one page in the browser: llama.cpp for chat, stable-diffusion.cpp for images, whisper.cpp for transcription and Kokoro for speech. Everything, including the models, sits in one folder, so you can carry it on a drive and run it offline.
A model is one big file (a GGUF). Before it can answer, the engine copies the whole file from the drive into memory. I think of it as a kitchen: the drive is the pantry down the hall, the RAM is the counter, and the chip is the cook. A pantry further away means a longer walk before cooking starts.
The setup
- Computer: Mac mini M4 with 16 GB of RAM, macOS 26.6.2
- Lane A: a SanDisk Extreme Portable SSD (500 GB, exFAT) over USB
- Lane B: the Mac mini’s internal SSD
- The same app at the same commit (3f0830f527d4), byte-identical model files, the same prompt
Every load was cold. Before each one I unmounted the drive and purged the file cache, so nothing came from memory.
Test 1: setup
First-run setup took 82 s on both drives, because most of it is downloading. Copying the installed app (24,751 small files) was where the USB drive lost: 41 to 43 s to the USB SSD and 8 to 14 s to the Mac’s disk.
Test 2: loading the model
Gemma 4 E4B (5 GB), five cold loads per drive. The USB SSD took 14.7 s every time and the Mac’s disk took 5.2 s. That gap is the drive’s read speed.
Test 3: speed once loaded
Once the model was in memory, the drive stopped mattering. Answers came at 25.4 tokens a second from the USB SSD and 25.7 from the Mac’s disk, and llama-bench gave about 30 for both. The same image took 32.2 s on each and came out byte for byte identical.
The race
With the Wi-Fi off, each lane did the whole job back to back: launch the app, load the model, answer a prompt, make an image, transcribe 30 seconds of audio and read an answer aloud. Five rounds, with the starting drive swapped each round.
| USB SSD | Mac’s disk | |
|---|---|---|
| Whole job (median) | 68.3 s | 55.7 s |
| Waiting on the drive | 21.2 s | 9.9 s |
| Doing the work | 45.1 s | 43.6 s |
Nearly all of the 12.6 s gap was the wait for the drive.
What fits in 16 GB of RAM
Then I climbed a ladder of Gemma 4 models, all at the same 4-bit quantization (Q4_K_M) and all loaded from the USB SSD.
| Model | File | Load | Tokens/s |
|---|---|---|---|
| Gemma 4 E2B | 3.1 GB | 9.7 s | 47.6 |
| Gemma 4 E4B | 5.0 GB | 14.7 s | 25.7 |
| Gemma 4 12B | 7.1 GB | 20.7 s | 11.8 |
| Gemma 4 31B | 18.3 GB | 7 min 43 s | crashed on the first answer |
The 31B is bigger than the Mac’s RAM. It pushed about 16 GB into swap and crashed before it answered. I ran it from both drives to see whether the drive matters again once the model doesn’t fit. From the USB SSD it loaded after 8 minutes and crashed on the first answer. From the Mac’s disk the app gave up after 6 minutes.
Windows
I plugged the same SSD into my Windows PC and it loaded the same model. Both GPU engines the app ships failed on my NVIDIA card, though, so it ran on the CPU at 3.7 tokens a second.
What I took from it
The drive sets how long you wait for the model to load. The RAM sets which models run and how fast they answer. If you’re buying, get enough memory for the model you want first, and after that any decent SSD will do. I didn’t test a cheap USB flash stick, which can be much slower than this SSD, and I didn’t test Linux.
Check your own drive with drive-check
I wrote drive-check for this. It’s free and open source. Point it at a drive and a model file, and it measures the drive’s real read speed, estimates how long the model takes to load, checks the model against your RAM and lists which Gemma sizes from my ladder fit your computer. You need Python 3.8 or newer and nothing else.
python3 drive_check.py /Volumes/MyDrive --model /Volumes/MyDrive/models/gemma-4-E4B-it-Q4_K_M.gguf
For my USB SSD it estimated 15 s, and the app took 14.7 s. On Windows, run py drive_check.py D:\ --model D:\models\gemma-4-E4B-it-Q4_K_M.gguf.
Run it before you buy a new drive or a bigger model, and tell me your load time in the video’s comments.
Code & video
- 📺 Full video: AI on a USB Drive: I Tested If It Actually Slows You Down
- âš¡ The Short: youtube.com/shorts/fdHC6X7lkLA
- 💻 drive-check: github.com/thecoderlog/drive-check
- 🧰 Portable Local Studio: github.com/techjarves/Portable-Local-Studio, and the video that started this, by Tech Jarves
- 📚 Engines: llama.cpp, stable-diffusion.cpp, whisper.cpp, Kokoro-82M
If this was useful, subscribe to The Coderlog on YouTube for the next test.