Home / Guides / Run AI Locally: Easy Setup

How to run AI locally with an easy setup (no command line)

Last updated: September 8, 2026

You want a ChatGPT-style assistant that runs on your own computer, does not need a subscription, and does not send what you type to anyone. You do not want to learn what a quantization is to get there. This guide is the short path: what your machine can handle, what to honestly expect, and a setup that makes the decisions for you.

Why people do this, and why they quit

The reasons are the same in every forum thread: privacy, no monthly bill, working offline, and the fun of owning it. The reason most people quit is the first session. They install a runtime, guess at a model, it runs slowly on the processor because the graphics card was never used, the answers are dumber than ChatGPT, and nobody warned them. All of that is avoidable if the setup picks the model for your hardware and tells you what to expect before you download anything.

Step 1: know which kind of machine you have

You do not need model names or memory sizes. There are three kinds of machine, and the only thing that matters is which one you own:

The graphics card decides almost everything. Below 8 GB of video memory, go small or accept processor speed. A recent Mac with 32 GB or more of memory counts as high-end because its memory is shared with the graphics chip. If you are not sure what you have, the setup below checks for you.

Step 2: set your expectations honestly

• A small model is not ChatGPT. It is a capable helper for questions, rewrites, and short summaries. It gets hard reasoning wrong more often.
• A mid model is close to a free chatbot tier. Good for most everyday work.
• A large model feels like a paid chatbot for most tasks. Not all of them.
• You will not save money versus paying per message unless you chat a lot. You do this for privacy and ownership.
• Raise the conversation memory on day one. Some tools default to a very short memory, which makes the model look dumber than it is. A good setup sizes it to your machine for you.

Step 3: the easy setup

This is the part most guides make hard. With LumaBrowser it is one download and one button:

1. Download the installer for your system. On Windows, click More info, then Run anyway if SmartScreen appears; the installer is not code-signed yet. On macOS, right-click the app and choose Open the first time. On Linux, run the one apt line on the download page first.
2. Open the app and choose Automatic Local Setup.
3. Answer three plain questions: what you will use it for, the slowest reply speed you would accept, and how much conversation memory you want. Say no to images unless you have a gaming graphics card.
4. Read the plan. It names the model it picked, why, how big the download is, and how much disk it needs. Click Start.
5. Wait. A few minutes on a laptop, up to half an hour for the biggest models. The download pauses and resumes; closing the app does not lose it.
6. Chat.

What happened underneath: the app read your memory, graphics card, and free disk; picked the largest model that runs comfortably on that hardware; fetched the matching engine build for your card (NVIDIA, AMD, Apple Silicon, or plain processor); sized the conversation memory to what fits; and started a local server. None of that needed you. The screen-by-screen version with install prompts is in the getting-started guide.

Step 4: the first session

The first reply takes a few extra seconds while the model loads. After that, try a rewrite ("make this email shorter and friendlier"), a summary (paste a page of text), and a plan ("three days in Lisbon, food and walking"). Then ask it "what are you bad at?" A local model will usually tell you: current events, long documents, and tricky maths, in roughly that order. Turn on web browsing when you need today's facts.

The five problems everyone hits, and the fixes

It is slow

No graphics card: slow typing speed is normal. Gaming card and still slow: the model is too big and spilling into system memory; pick a smaller one in the Setup tab, and update your graphics driver.

It is dumber than I hoped

Check which tier you are in above. If you have a gaming card and got a small model, the setup may have played safe; try the next size up and watch the speed.

The download stalled or the disk is full

Downloads resume. Models are big; keep 10, 30, or 80 GB free depending on your tier, and move the models folder to another drive from Settings if C: is tight.

Windows or antivirus blocked the installer

New unsigned installers get flagged. More info, Run anyway; and check the file size and checksum on the download page if you want to be sure.

My laptop got hot and loud

While a reply is being written, yes. Between replies it idles. Plug in for long sessions.

If you want to go deeper

Everything above is the default. If you would rather understand what the setup decided, how to run an LLM locally explains memory, model files, and engine builds from scratch, and the advanced page exposes every knob. Comparing tools first? LumaBrowser vs LM Studio, Jan, and Ollama.