The free guide

This is the whole method, given away. Everything below uses free, open-source software, and if you follow it you will have a working private AI on a USB drive without paying us anything. We sell the packaged version because most people would rather not do this — not because the information is secret.

What you need before you start

Item	Minimum	Comfortable	Why it matters
USB drive	8 GB, USB 3.0	32 GB, USB 3.2 or SSD	Read speed decides how long the model takes to load. Capacity is rarely the limit.
Host RAM	8 GB	16 GB	The model is loaded into the computer's RAM. This is the real constraint, not the drive.
Windows	10 (64-bit)	11	The runtime below ships a 64-bit Windows build.
Free disk on host	none required	—	Everything can live on the drive if you point the runtime at it.

Breakdown of an 8 GB drive: model weights about 2.5 GB, the runtime about 30 MB, the launcher
and interface 0.2 MB, leaving roughly 5.5 GB free. — To scale. The model weights are almost the entire package — and they are the part that is free.

Step 1 — Choose a model you are allowed to use

This step is skipped by almost every guide on the internet, and it is the one that matters most if you intend to do anything commercial. Open weights are not the same as an open licence. Some well-known models are published under terms that permit research and evaluation only.

Model	Licence	Roughly	Commercial use
Qwen3 4B	Apache 2.0	2.5 GB	Yes, with attribution
Phi-4 mini	MIT	2.5 GB	Yes, with attribution
SmolLM2 135M	Apache 2.0	0.27 GB	Yes, with attribution
Qwen2.5 3B	Qwen Research Licence	1.9 GB	No — non-commercial only
Llama 3.2 3B	Llama Community Licence	2.0 GB	Conditional — notice and naming rules apply
Gemma 3 1B	Gemma Terms of Use	0.82 GB	Conditional — restrictions pass downstream

For personal use any of these is fine. If you are building something you intend to sell, stay in the top three. Read the licence yourself before you rely on this table — you can print the exact terms your copy shipped with:

ollama show qwen3:4b --license

We use Qwen3 4B in the packaged build for exactly this reason. It is not the smallest option; it is the most capable one we are clearly entitled to redistribute.

Step 2 — Put the runtime on the drive

You need something that can load a model file and answer prompts. Ollama is the simplest on Windows and is MIT-licensed. Install it normally first — you are going to point it at the drive afterwards.

Install Ollama from its official site and let it finish.
Make a folder on the USB drive, for example D:\ai\models.
Tell the runtime to keep model files there by setting OLLAMA_MODELS to that folder before you launch it. This is the whole trick: the runtime reads from the drive instead of from your user profile.
Pull the model with ollama pull qwen3:4b. It downloads once, onto the drive.

Step 3 — Prove it is actually offline

Three steps: cut the network, ask it a question, and if it answers then the local machine did
the work. — Run this before you trust any claim on this page, including ours.

Do not take anyone's word for this, including ours. Turn off wifi, unplug the network cable, and ask it a question. If it answers, the computer in front of you did the work. If it fails, something in your setup is still reaching for a server.

A second check worth doing: watch the network while it answers. A local model generates no outbound traffic during inference.

Step 4 — Make it start from the drive

Write a small batch file on the drive that sets OLLAMA_MODELS, starts the runtime, and opens whatever interface you want to use. Double-clicking that file is your launcher.

On autorun: Windows deliberately disabled automatic execution from USB drives, because that is precisely how a generation of malware spread. Do not try to defeat it, and be wary of any product that offers to. A launcher the user double-clicks is the correct design, and it is what we ship.

What you will notice

The first answer is slow. The model is being read off the drive into RAM. On a USB 3.0 stick expect tens of seconds. Later answers in the same session are far quicker because the model is already resident.
It is not a frontier model. A 4-billion-parameter model that fits on a keyring will not match a hosted system running on datacentre GPUs. It is genuinely useful for drafting, summarising, explaining and rewriting.
Cheap drives are the usual disappointment. A slow USB 2.0 stick can triple your load time. This is the single upgrade that helps most.

So what are we selling?

The steps above, done for you and tested: a launcher that finds the runtime whether or not it is installed, hardware detection so it picks a model your machine can actually hold, a real interface instead of a terminal, the licence paperwork already sorted, and a year of updates. $9.99 once, then $9.99 a year if you want to keep getting updates. If you would rather spend an afternoon than ten dollars, everything you need is on this page and we are glad to have helped.

Questions about this page

Is this guide really complete, or is something held back?

It is complete. Every step needed to get a local model running from a USB drive is on this page, including the licensing step most guides omit. What we sell is assembly and testing, not withheld information.

Can I sell what I build by following this?

Only if the model licence allows it. Qwen3 4B, Phi-4 mini and SmolLM2 are permissively licensed and permit commercial use with attribution. Qwen2.5 is research-only. Llama and Gemma carry conditions you must read and satisfy. Check the licence of your exact model before you sell anything.

Why does the first answer take so long?

Because the model file is being read from the USB drive into the computer's memory before it can answer. That is a one-time cost per session and it is dominated by the read speed of the drive. Later answers are much faster.

Will this work on a Mac or on Linux?

The concepts all carry over and the runtime exists for both. This guide gives Windows specifics because the packaged product is Windows-only today. We have not tested the packaged build on macOS or Linux, so we do not claim it works there.

Do I need a graphics card?

No. A model this size runs on the CPU. A GPU makes it faster but is not required, and on a portable setup you should assume you will not have one.