How to Run AI on Your Own Computer: A Beginner's Guide

When you use ChatGPT or Claude, your messages travel to the company's servers and the answer comes back. Local AI works differently: the model is a file on your computer, and everything happens on your own machine. Your prompts never leave it, it works offline, and there's no subscription or usage limit.

A few years ago this took serious technical skill. Today, free apps like Ollama and LM Studio make it a ten-minute job. This guide explains what you need, how to get started with either app, and which model to download first.

Is local AI right for you?

Local AI is a great fit when:

• Privacy matters — you want to work with personal notes, client documents or code without sending them to anyone.
• You want no limits or bills — use it as much as you like once the model is downloaded.
• You're often offline, or just curious how these models work.

The honest trade-off: the models a normal computer can run are smaller than the top cloud models. They're good at everyday writing, summarizing, brainstorming and simple coding, but they fall behind the best paid assistants on hard reasoning and specialist knowledge. Many people use both — local for private tasks, cloud for the hardest ones.

What hardware do you need?

The key number is memory. The model has to fit into your computer's memory (RAM), or into your graphics card's memory if you have one, with room left for everything else you're running. Bigger models are smarter but need more memory.

A few rules of thumb:

• A graphics card makes it much faster. Ollama supports NVIDIA and AMD graphics cards, and Apple Silicon Macs (M-series) through Metal. Without one, small models still run on the processor — just more slowly.
• Apple Silicon Macs are well suited, because their memory is shared between the processor and graphics, so a 16 GB Mac can use most of that for a model.
• Leave some disk space. Each model is a multi-gigabyte download, and you'll probably try a few.

Option 1: Ollama — the simplest start

Ollama is a free app for Windows, macOS and Linux that downloads and runs models for you. It includes a simple chat window, and it also works from the command line.

1. Install it from ollama.com. On Windows and Mac, run the installer like any other app.
2. Start a model. Open the Ollama app and pick a model in the chat window, or open a terminal and type:

   ollama run qwen3.5:4b

   The first time, this downloads the model (a few gigabytes). Then you can type your message right away. If you have 16 GB of memory or more, try ollama run gemma4 for a stronger medium-sized model.
3. Chat. Ask anything. In the terminal, type /bye to leave.

Three more commands cover most needs:

ollama list              # show the models you've downloaded
ollama ps                # show which models are loaded right now
ollama rm qwen3.5:4b     # delete a model to free up disk space

One thing to know: Ollama also offers cloud models, which run on Ollama's servers instead of your computer — they have "cloud" in their name, and heavier use requires a paid plan. If privacy is your reason for going local, stick to models without "cloud" in the name. You can also switch cloud features off completely by setting the environment variable OLLAMA_NO_CLOUD=1.

For developers, Ollama runs a local API at http://localhost:11434, so your own scripts and apps can use the model.

Option 2: LM Studio — point and click

LM Studio is a free desktop app — including for work use — with a friendly interface and no command line needed.

1. Install LM Studio from lmstudio.ai.
2. Find a model with the built-in model search. LM Studio shows how big each download is and warns you when one is likely too large for your machine.
3. Download it, load it and chat in the app's chat window, much like ChatGPT.

LM Studio can also run a local server that speaks the same API format as OpenAI's, which makes it easy to point existing tools at your local model.

Ollama or LM Studio? Pick LM Studio if you prefer browsing and clicking; pick Ollama if you like a quick command or want to build on its API. Both are free, and you can install both.

Which model should you start with?

Model libraries change fast, but these families are well established and available in both apps:

• Qwen 3.5 (Alibaba) — comes in many sizes, from tiny to huge, so it's easy to find one that fits your machine.
• Gemma 4 (Google) — capable small and medium models that suit laptops.
• Mistral — efficient models from the French lab Mistral AI, strong in European languages.
• Llama (Meta) — the smaller Llama 3 models (for example llama3.1:8b) run well at home. Meta's newest Llama 4 models are much larger, with over 100 billion parameters in total, so they're out of reach for most personal computers.
• For coding, try a code-tuned model such as qwen2.5-coder.

Start small. If the answers are too basic and your computer has room, move up a size. If it's too slow, move down.

Before you use a model for business, check its license on the model page. Licenses differ: Llama, for example, comes with Meta's own license terms rather than a standard open-source license.

Beyond chat: images and transcription

• Stable Diffusion generates images locally with no limits. It's more demanding than chat models, so a dedicated graphics card helps a lot. It's free for individuals and businesses earning under $1M a year. New to AI images? Start with our guide on generating images with AI for free.
• Whisper is OpenAI's open-source speech-to-text model. It transcribes recordings in 99 languages, entirely on your machine.

Local vs cloud at a glance

If something goes wrong

• Very slow answers usually mean the model is too big for your memory. Try a smaller size, and close memory-hungry apps like browsers with many tabs.
• The model won't load — same cause. Check the download size against your free memory.
• Strange or repetitive answers are more common with very small models. Move up one size if you can.

Stay safe

• Download models from the apps' own libraries or well-known sources, not from random links.
• Keep the local server local. Ollama listens only on your own computer by default. Don't open its port (11434) to the internet unless you know how to secure it — anyone who can reach it can use your machine's AI.

Want to compare all the options? See our list of the best open-source and self-hosted AI.