Guide · step by step · any OS

Run AI on your own computer

Install a local AI runner, download a model, and chat with it offline. Steps for Windows, macOS and Linux, with the commands you need and the common problems.

FreeWindows · macOS · Linux

Quick path: install Ollama, open a terminal, run ollama run llama3.2, and type a question. That is the whole idea; the rest of this guide explains each step and what to do when something goes wrong.

Jump to: What you need · Windows · macOS · Linux · First models · Commands · Local API · Pick by RAM · Alternatives · Problems · Uninstall

What you need

A computer from roughly the last 8 years. Any modern 64-bit Windows, macOS or Linux works for small models.
RAM is the main limit. 8 GB runs small models (about 3-4 billion parameters); 16 GB runs 7-8B models comfortably; 32 GB opens up 14B and some larger ones.
A graphics card (GPU) or an Apple Silicon Mac makes answers much faster, but it is optional; a CPU alone works, just slower.
Free disk space: a small model is 2-5 GB, a mid-size one 5-15 GB. Keep 20 GB free to start.
An internet connection once, to download the program and the model. After that it works offline.

Not sure what your machine can handle? Use the sibling tool Can my PC run this AI model? to see which models fit before you download anything.

Step 1. Install Ollama

Ollama is a free program that downloads models and runs them, with a simple command line and a local API. Pick your system; all three sections are shown so you can read them without JavaScript.

Windows

Download the installer from ollama.com/download and run it (no administrator rights are normally needed).
Or, from a terminal (PowerShell or Command Prompt):
```
winget install Ollama.Ollama
```
Ollama starts in the background (look for the llama icon in the system tray). Open PowerShell and check:
```
ollama --version
```

Windows Defender or SmartScreen may ask for confirmation the first time. Only download from the official site.

macOS

Download the app from ollama.com/download, open the downloaded file and move Ollama to Applications. Start it once.
Or, with Homebrew:
```
brew install ollama
```
With the Homebrew version you start the server yourself in one terminal window:
```
ollama serve
```
and use the commands in another window.
Check:
```
ollama --version
```

Apple Silicon Macs (M1 and later) use the GPU automatically and share memory between CPU and GPU, so RAM size decides what fits.

Linux

Run the official install script (read it first if you like; it needs sudo and installs a system service):
```
curl -fsSL https://ollama.com/install.sh | sh
```

Check that the service is running:

ollama --version
systemctl status ollama

Without systemd, start the server manually with ollama serve in a terminal.

NVIDIA and AMD graphics cards need their normal drivers installed for GPU use. If no supported GPU is found, Ollama falls back to the CPU.

Step 2. Run your first model

The first ollama run of a model downloads it, then opens a chat. Type your question, press Enter; type /bye to leave.

ollama run llama3.2

Command	Good for	Rough download
`ollama run llama3.2`	Small and quick; a fine first test on almost any machine (3B default)	about 2 GB
`ollama run gemma3:4b`	Small, capable, good for everyday questions	about 3 GB
`ollama run qwen3:8b`	Stronger general model for 16 GB machines	about 5 GB
`ollama run gpt-oss:20b`	Larger open-weight model; plan for about 16 GB of free memory	about 13-14 GB

Sizes are approximate and change as models are updated; the model page on ollama.com shows the exact number. Model names and tags change over time, so verify a name there if a pull fails.

Everyday commands

Command	What it does
`ollama list`	Show the models you have downloaded and their sizes
`ollama pull qwen3:8b`	Download (or update) a model without starting a chat
`ollama rm qwen3:8b`	Delete a model to free disk space
`ollama ps`	Show which models are loaded in memory right now
`ollama show llama3.2`	Show details of a model (size, parameters, template)
`ollama stop llama3.2`	Unload a model from memory
`ollama run MODEL "Your question"`	Ask one question and print the answer, no chat window

Models stay in memory for a few minutes after your last request, then unload on their own.

Where models are stored

System	Default location
Windows	`C:\Users\<you>\.ollama\models`
macOS	`~/.ollama/models`
Linux (service install)	`/usr/share/ollama/.ollama/models`

To keep models on another drive, set the OLLAMA_MODELS environment variable to a folder there and restart Ollama. On Linux with the service, the variable must be set for the service (for example with sudo systemctl edit ollama) and the ollama user needs write access to the folder. Verify against the Ollama FAQ for your version.

The local API

While Ollama runs, it listens on http://localhost:11434, only on your own computer by default. Other programs (editors, note apps, scripts) can use it. Quick test:

curl http://localhost:11434/api/generate -d '{"model":"llama3.2","prompt":"Say hello in five words.","stream":false}'

curl http://localhost:11434/api/tags

The second command lists your downloaded models. On Windows PowerShell, use curl.exe (plain curl is an alias for something else) or run the commands from Command Prompt. Many tools can also talk to Ollama through its OpenAI-style endpoint at http://localhost:11434/v1/.

Do not expose this port to the internet or your network without understanding the consequences: the API has no password.

Which model for how much RAM

A rule of thumb for the common 4-bit versions Ollama downloads: the model file is about 0.6 GB per billion parameters, and you need that plus a few GB for the conversation and your system. Unified-memory Macs and GPUs: use the GPU memory (VRAM) instead where it is separate.

Memory available	Sensible model size	Examples
4 GB	1-2B	`llama3.2:1b`
8 GB	3-4B	`llama3.2`, `gemma3:4b`
16 GB	7-8B (14B is tight)	`qwen3:8b`, `gpt-oss:20b` only if memory is free
32 GB	up to about 14-20B comfortably	`qwen3:14b`, `gpt-oss:20b`
64 GB and more	30B and up	see the tool linked above

These are rough planning numbers, not guarantees. The model-fit calculator gives a closer estimate for your hardware.

Alternatives

LM Studio (lmstudio.ai): a desktop app with a model browser and chat window, good if you prefer clicking to typing commands.
llama.cpp (github.com/ggml-org/llama.cpp): the engine many tools are built on. More control and more setup; runs GGUF model files with llama-cli and llama-server.
LocalAI (localai.io): a self-hosted server with an OpenAI-compatible API, often run in Docker.

Common problems

Out of memory, or the computer freezes

The model is too big for your RAM or GPU memory. Pick a smaller one (see the table), close other heavy programs, or use a shorter context. ollama ps shows what is loaded; ollama stop MODEL frees it.

Very slow answers

Usually the model runs on the CPU, or part of it spilled out of GPU memory into system RAM. Try a smaller model or a smaller context. ollama ps shows how much sits on the GPU versus CPU. The first answer after loading is slower than later ones.

The GPU is not used

Update your graphics driver (NVIDIA or AMD), restart Ollama, and check again with ollama ps. On Linux, check the service log with journalctl -u ollama. Not every older card is supported; check ollama.com and its GitHub page for the supported list.

"connection refused" on port 11434

The server is not running. Start the app (Windows, macOS), run ollama serve, or on Linux sudo systemctl start ollama.

Download stops or fails

Run the same ollama pull again; it continues where it stopped. Check free disk space.

Uninstall

Windows: Settings, Apps, Installed apps, Ollama, Uninstall. Delete the models folder (.ollama in your user folder) to reclaim disk space.
macOS: quit Ollama, delete it from Applications (or brew uninstall ollama), and remove ~/.ollama to delete models.
Linux: stop and disable the service, then remove the files the installer created. The Ollama docs list the exact steps; in outline:
```
sudo systemctl stop ollama
sudo systemctl disable ollama
sudo rm /etc/systemd/system/ollama.service
sudo rm "$(which ollama)"
sudo rm -r /usr/share/ollama
```
Verify against the current docs, and read each line before running it.

Check ollama.com for current instructions. Install methods, model names and sizes change. Commands in this guide were last checked on 2026-11-07 against what we know of Ollama; if something differs, the official docs win.