Run an AI directly on your computer, with no account, no subscription, and without sending your text over the internet: it is possible, free, and simpler than you might think. This guide walks you through it from start to finish — check your machine, choose the tool, download a first model, and run the test that proves everything stays at home.
What you need before you start: a reasonably recent computer, an internet connection for the initial downloads (after that, you no longer need it), and between 10 and 50 GB of free disk space — each downloaded model takes up several GB.
Step 1 — Check your machine
The only thing that really matters at the start is your computer’s memory (RAM):
8 GB of RAM: you can run small models (3 to 4 billion parameters). Answers will be slower and simpler, but it works.
16 GB of RAM: the real comfortable starting point. Models with 7 to 8 billion parameters — the most common and best balanced — run well.
32 GB of RAM (or a graphics card with a lot of memory): the large models with 14 to 32 billion parameters, which are noticeably better, become realistic.
Two useful clarifications:
A graphics card is not required for chatting: the processor is enough, just slower. It becomes almost essential only for generating images or videos.
Macs with an Apple Silicon chip (M1, M2, M3, M4…) are especially comfortable, because their memory serves both the system and graphics computation. An older Intel Mac can still manage with Jan or GPT4All and a small model.
Where do you look? On Windows: Settings → System → About (“Installed RAM”). On a Mac: Apple menu → About This Mac.
Step 2 — Choose your entry point based on your profile
No need to try everything: the tools are similar, so choose the one that matches your situation. All of them are free for the local use described here.
“I want the simplest option” → LM Studio. The model catalog is built in, the interface is familiar, and the app is free even at work. The only condition: an Apple Silicon Mac or a recent PC. See the LM Studio profile
“I want open source, and true offline use” → Jan. Free and open source, it runs even on an Intel Mac, and the conversation continues with Wi-Fi turned off. See the Jan profile
“My computer is modest” → GPT4All. Lightweight models designed for the processor alone. See the GPT4All profile
“It’s my documents I care about” → AnythingLLM. You drop your files into a workspace and query them, without them ever leaving the machine. See the AnythingLLM profile
“I want to compare before I choose” → Msty. It asks the same question of several models and displays the answers side by side. See the Msty profile
Step 3 — Download a sensible first model
The app you chose will offer you a catalog: these are the “models,” the brains you can plug in (Llama, Qwen, DeepSeek, Gemma, Mistral…).
Start with a model of 4 to 8 billion parameters (often written “4B” or “8B”), in a “quantized” version if the app offers one: it is the standard lightweight format, made for personal computers. The file is about 3 to 6 GB.
Resist the biggest model of the moment: a 32-billion-parameter model on a machine with 16 GB of RAM will mostly be… slow. You can always download a second one later.
You download it only once. After that, the model lives on your disk and no longer needs the internet.
Step 4 — Run the offline test
This is the moment of truth — and the whole point of this path:
1. Launch your app and open a conversation with the downloaded model.
2. Turn off Wi-Fi (or enable airplane mode).
3. Ask a real question: a text to rephrase, an idea to expand, a summary.
The answer still arrives. Nothing was sent to a server: your words never left your computer. That is the concrete difference from ChatGPT or Gemini, and why local AI is the most private solution — for a sensitive draft, a client file, or a personal journal, there is nothing better.
Step 5 — The three normal disappointments (so you don’t give up too early)
Almost everyone runs into them. They are normal; it does not mean you installed something wrong:
1. “It’s slower than ChatGPT.” Yes: your computer is doing alone the work that entire data centers do online. A smaller model will answer faster; a 4-billion-parameter model is often the right everyday compromise.
2. “The answers are a notch below.” Also true: the free models you can run at home are less powerful than the best online models, especially for long reasoning and specialized topics. And like all models, they can be confidently wrong: check what matters, and for your documents, rely on the source passages when the tool cites them (GPT4All and AnythingLLM do).
3. “The first download is huge.” Unavoidable: several GB for the model, sometimes as much for the app. It is a one-time cost — plan it on a good connection, not on a limited mobile plan.
And the guardrails to keep in mind:
Download apps only from the publisher’s official website (the addresses are in each Directory profile). Never from a third-party “free download” site: that is the classic gateway for malware.
“Local” only protects what stays local. If, in Jan, Msty, or AnythingLLM, you plug in an online model (through a cloud provider’s API key), your text goes back out over the internet. For privacy, stick to models labeled “local.”
Back up what matters, as usual: your local conversations and documents live on your disk — if the disk fails, they go with it.
Step 6 — Going further
Once your first model is running, depending on what you feel like:
Agents that work at your home: Ollama is the reference engine for connecting interfaces and agents to your local models — a bit more technical, and the logical next step after this guide. In the same vein, OpenClaw installs an always-on personal assistant in your messaging apps.
Image, voice, and video applications: Pinokio installs in one click the open-source AI applications that normally require the command line. Important guardrail: a “one-click” launcher runs code on your computer — only install applications from the official catalog or from identifiable authors, never a script found at random on a forum.
A home web interface (step 2, for the more comfortable): Open WebUI plugs into Ollama and is self-hosted, but installing it goes through Docker and the command line. We mention it here without a beginner profile.