Run an AI Model on Your Own PC, With Your Kid

Everything in this guide is free, runs on the computer already in your house, and sends nothing to anyone. By the end of it a parent and a teenager will have a capable AI assistant on the family PC, with the parent holding the admin account and able to read every conversation. It is the supervised version of what The Gate Has a Hole in It describes a fifteen-year-old doing alone in a bedroom. Do it together instead.

What you need

A desktop or laptop with a dedicated graphics card is the comfortable path. The number that matters is the card's memory, VRAM, not the computer's RAM. Here is what fits where, using Q4 quantised models, which is the normal way to run them at home:

8 GB VRAM (RTX 3060, 4060, most gaming laptops): an 8-billion-parameter model such as Qwen3 8B. Fast, competent, a little shallow on hard reasoning.
12 GB VRAM (RTX 3060 12 GB, 4070): a 14-billion-parameter model. This is the biggest single jump in quality for the money.
16 GB VRAM (RTX 4080, 5070 Ti): Gemma 3 27B or Mistral Small 24B, with a shorter context window.
24 GB VRAM or more (RTX 3090, 4090, 5090): Qwen3.6 27B, the model named in the article, at about 17 GB. This is the one that does serious work.
Apple Silicon Mac: unified memory counts as VRAM. 16 GB runs the 8B and 14B models; 32 GB or more runs Qwen3.6 27B.
No graphics card: a 3 to 4-billion-parameter model will run on the CPU with 16 GB of RAM. Slow, but it works, and it is enough to learn the mechanics.

One honest note. The other model named in the article, DeepSeek V4 Flash, is 284 billion parameters and needs more than 110 GB of memory even at its smallest usable size. It is open-weight, and that matters for the argument, but it is not a bedroom model. The Qwen family is.

Step 1: Install Ollama

Ollama is the program that downloads and runs the models. Go to ollama.com/download, pick Windows, Mac or Linux, and run the installer. On Windows it adds an icon to the system tray; on Mac, to the menu bar. That is the whole install.

Open a terminal (Windows: search for PowerShell; Mac: search for Terminal) and type:

ollama run qwen3:8b

It downloads about five gigabytes and then gives you a prompt. Ask it something. You are now running a language model on your own hardware. Type /bye to leave.

Pick the model for your card from the list above. The tags are qwen3:8b, qwen3:14b, gemma3:27b, mistral-small and qwen3.6:27b. For a CPU-only machine, qwen3:4b. You can have several installed and switch between them.

Step 2: Put a proper chat window on it

The terminal is fine for a test. For daily use you want Open WebUI, a free web interface that runs on the same machine and talks to Ollama. It looks and behaves like the commercial chat sites, keeps every conversation, and, crucially, has user accounts.

The simplest install uses Docker. Install Docker Desktop, start it, then in the terminal:

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

Give it a minute, then open http://localhost:3000 in a browser. The first account created is the administrator. The parent creates that account. This is the step that makes everything else work.

If Docker is more than you want, Open WebUI also installs with a single pip install open-webui on a machine that has Python, and starts with open-webui serve. Either way the address is the same.

Step 3: Give your teenager their own login

In Open WebUI, as the admin, open the Admin Panel from the menu in the bottom left. Under Users, add a user for your child with their own email and password, role user. From now on they sign in as themselves, you sign in as yourself, and the two are separate.

Two settings in the Admin Panel are worth setting now. Under Settings, turn off new sign-ups, so nobody else can create an account on the box. Under Models, you can restrict which models the user role can see, which is useful if you have installed a large model you would rather they ask before using.

Step 4: Keep the log visible

This is the part that turns a bedroom setup into a supervised one, and it is the part the VEIL Standard's Guardian tier is built around: the parent can read the session log.

In the Admin Panel, under Users, click your child's name. Their chat list is there. You can open any conversation and read it in full. Nothing is hidden from the admin account, and nothing leaves the machine, so the only people who will ever see those conversations are the two of you. Say so, out loud, before they start. The deal is simple: they get a tutor that never sleeps and never reports them to anyone; you get to read the transcript.

Most families will find they rarely look. The point is not surveillance, it is that the option exists, the child knows it exists, and the record is in the house rather than on a server in Virginia. That is the entire difference between this and the thing the codes are trying to ban.

Step 5: Set the ceiling

A local model has had its safety training, but it has no age gate and no account to be banned from. The ceiling is you. A few practical settings help:

Stick to the official model tags. There are modified versions on the internet with the safety training stripped out, usually labelled "uncensored" or "abliterated". Do not install those on a machine a child uses.
In Open WebUI, under Workspace, you can give a model a system prompt. Something like "You are a tutor for a fifteen-year-old. Explain, show working, and decline anything you would not say in front of their parent" is enough. The model will mostly honour it, and the transcript will show where it did not.
Keep the machine in a shared room if the child is young. For an older teenager, the transcript is the room.

Step 6: Install the VEIL Guardian

Steps 3 to 5 are the manual version of what the VEIL Standard's Guardian tier describes. This step installs it properly. VEIL Guardian is a free filter for Open WebUI: once it is in, every message from your child's account carries a Tier G envelope, the content ceilings are locked to their age profile, the six minor-safety obligations are in front of the model on every turn, crisis language puts Kids Helpline and Lifeline in front of the model, and an audit line is written per turn. If the envelope is ever missing or broken, the child's account is paused until you fix it.

Open veil_guardian.py on GitHub, click Raw, select all, copy.
In Open WebUI as admin: Admin Panel → Functions → New Function. Paste over the editor. Save.
Switch the new VEIL Guardian (Tier G) row on, then open its ⋮ menu and switch Global on.
Click the ⚙ Valves icon. In envelope_json set creator_id to your child's login email, guardian.creator_id to yours, and minor_profile to VEIL-U13, VEIL-13 or VEIL-16. Set guardian_label to what they call you. Save.
Sign in as your child and ask anything. A small line VEIL Tier G · VEIL-16 · envelope … appears on the reply. Ask it "is this private?" and it will say the session is logged and visible to you.

The full instructions and the valve reference are in the repository. The envelope you filled in is a real VEIL v1.2 context object; the same file works with any engine that takes a system prompt.

What it is like to use

On a 24 GB card, Qwen3.6 27B answers at reading speed. It will explain the firing order of a four-cylinder engine with the crank geometry, walk through a compile error line by line, set out a phase diagram, check the arithmetic on a print tolerance and translate a Latin inscription. It does not browse the web, so it is sometimes out of date and sometimes confidently wrong, and learning to notice that is itself a skill worth having. On an 8 GB card the 8B model does the same things with less depth and more mistakes; it is still a better tutor than a worksheet.

Both of you will learn something in the first weekend. The teenager learns what a model is, what quantisation does, why VRAM matters and how a local web service works. The parent learns that the thing the headlines describe as a danger is, on the family PC with the log open, a very patient teacher.

Where this sits

With Step 6 done, this is a Guardian-tier VEIL session in the full sense: a verified parent, a child profile with ceilings locked to their age, a session log the guardian can read, and an engine that discloses it is an AI and keeps no secrets from the parent. Nothing in the eSafety codes gives a family a way to consent to this. A free download does. If the law ever catches up, the gate it will need is already drawn at r8rly.org/veil, and the family that has run this guide will already be living inside it.

Companion pieces: The Gate Has a Hole in It and The Permission Slip Generation.