Follow these in order and you'll have a private AI running on your own computer in about 20 minutes. No coding, no command line, nothing to pay. (Want a zero-setup taste first? Go to z.ai and chat with the GLM-5.2 flagship free in your browser — just remember it's a hosted Chinese service, so keep confidential material out of it. The real, private version is below.)
Step 1 — Check how much memory (RAM) your computer has
This one number decides which models you can run, so start here:
- Windows: Settings → System → About → look for "Installed RAM."
- Mac: Apple menu → About This Mac → look for "Memory."
The rule of thumb: you need about 1 GB of RAM per billion "parameters" in a model's name, plus headroom for the computer itself. So 8 GB runs a 4-billion (4B) model comfortably; 16 GB handles up to about a 12B; 32 GB opens the bigger 27B-class models. Write your number down — you'll use it in Step 3.
Step 2 — Install LM Studio
Go to lmstudio.ai, download the version for your computer (Mac or Windows), and install it like any other app. It's free, and it gives you a friendly point-and-click screen — no typing commands. (If you happen to like the command line, Ollama does the identical job.)
Step 3 — Pick a model that fits your RAM
Open LM Studio's model search and choose one from the tier that matches the number from Step 1. Each is labeled with its maker and country so you can choose on capability and comfort:
8 GB of RAM (most everyday laptops):
- Gemma 3 4B — Google (US)
- Llama 3.2 3B — Meta (US)
- Phi-4-mini — Microsoft (US)
16 GB of RAM (the sweet spot):
- Gemma 3 12B — Google (US)
- Mistral Small — Mistral (France)
- Llama 3 8B — Meta (US)
- GLM-4 9B — Zhipu (China)
- Qwen 7B — Alibaba (China)
32 GB of RAM or more:
- Gemma 3 27B — Google (US)
- Qwen 27B — Alibaba (China)
Not sure? On a normal laptop, Gemma 3 4B (Google) is the easiest, most reliable first pick of the whole list.
Step 4 — Download the model
In LM Studio, type the model's name into the search box, click it, and hit download. The file is large (a few gigabytes), so give it a few minutes on a decent connection. You only do this once — after that it lives on your computer.
Step 5 — Start a chat, then turn off your wifi
Open a new chat with the model you just downloaded. Now turn off your wifi — and notice it still answers. That's the entire point: no subscription, no internet, no data leaving your machine.
Step 6 — Give it real work
Try the things you'd actually use it for: paste a meeting transcript and ask for a summary, hand it a budget table and ask what changed, or have it draft a public-records request. A small model is weaker than ChatGPT — but it's private, free, and yours.
If you'd rather not run a Chinese model
Completely fair. GLM and Qwen are excellent, but the menu above is full of alternatives: Google's Gemma (US), Meta's Llama (US), and Mistral (France) are all open, free, and install the exact same way. One honest note: once you're running any model locally with the internet off, the country of origin stops mattering for privacy — a model on your machine can't send your data anywhere. But if you'd simply rather not, you've got great non-Chinese choices.
One security rule
Only download models from the official source — the ones built into LM Studio's or Ollama's own libraries, or a maker's verified account on HuggingFace. Open model files can be tampered with (researchers have built "poisoned" models that hide a backdoor and pass every normal test), and a random re-upload from an account you don't recognize is exactly where a bad actor would hide one. Stick to the official library and you're fine.
That's the whole thing. Twenty minutes, and you've run a private AI that owes nothing to anyone — a small taste of what your whole newsroom may be running by this time next year.
