How to Install Ollama — Complete Beginner's Guide (2026)

Ollama is the simplest way to run large language models on your own computer. It handles the download, the model format, and GPU acceleration for you, and exposes one command — ollama run — that gets you a working private model. This guide installs it on all three operating systems using the official commands, runs a first model, and fixes the errors people hit most often.

If you have not picked hardware yet, a machine with 8GB of RAM will run small models on CPU, but a dedicated NVIDIA GPU with 8GB or more of VRAM is where local AI starts to feel fast. The VRAM requirements guide maps model sizes to memory, and the best GPU for local LLMs guide turns that into a purchase.

What You Need Before Starting

Minimum to get it working:

• Any computer from roughly the last five years
• 8GB RAM (enough for 7B–8B models on CPU)
• 10GB free disk space for your first model
• An internet connection for the one-time download

Recommended for a comfortable experience:

• A dedicated NVIDIA GPU with 8GB+ VRAM, or an Apple Silicon Mac (M1 or newer)
• 16GB RAM
• An SSD — model loading is dramatically faster than on a hard drive

Ollama runs on CPU by default and uses a compatible GPU automatically when one is present. You do not need to configure anything to get CPU inference working; GPU acceleration is detected at runtime.

Step-By-Step Setup

macOS

1. Go to ollama.com/download and download the macOS build, or install with Homebrew: brew install ollama.
2. If you downloaded the app, unzip it and drag Ollama.app into your Applications folder, then launch it once so it can register the background service.
3. Open Terminal and confirm the install:

ollama --version

Apple Silicon Macs (M1/M2/M3/M4) use the GPU and unified memory automatically — no extra drivers needed.

Windows

1. Download the installer from ollama.com/download (an OllamaSetup.exe file).
2. Double-click the .exe and follow the prompts. Ollama installs as a background service and adds the ollama command to your PATH.
3. Open PowerShell or Command Prompt and verify:

ollama --version

On Windows, an NVIDIA GPU with a current driver is used automatically. You do not need to install the CUDA toolkit separately for inference.

Linux

The official one-line install handles the binary and the systemd service:

curl -fsSL https://ollama.com/install.sh | sh

Prefer a manual install? Extract the official tarball instead:

curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst | sudo tar x -C /usr
ollama serve

Use the ollama-linux-arm64.tar.zst package on ARM64 systems. To run Ollama as a background service with systemd:

sudo systemctl daemon-reload
sudo systemctl enable ollama
sudo systemctl start ollama
sudo systemctl status ollama

NVIDIA on Linux: install your distribution’s NVIDIA driver and, if needed, CUDA, then confirm the GPU is visible with nvidia-smi. AMD on Linux: add the ROCm package (ollama-linux-amd64-rocm.tar.zst) and install ROCm v7 separately. ROCm support is real in 2026 but still less plug-and-play than CUDA — see the AMD notes in our best GPU guide.

Run your first model

With Ollama installed, download and run a model in one command:

ollama run llama3.3:8b

The first run pulls the model (a few GB), then drops you into an interactive chat. Type a question, press Enter, and the response is generated entirely on your hardware. Type /bye to exit.

If you want something smaller to start, ollama run llama3.2:3b downloads faster and runs comfortably on 8GB machines.

The Commands You Will Actually Use

Command	What it does
ollama run <model>	Download (if needed) and chat with a model
ollama pull <model>	Download a model without starting a chat
ollama list	Show models you have downloaded
ollama ps	Show models currently loaded in memory
ollama stop <model>	Unload a running model from memory
ollama rm <model>	Delete a model to free disk space

Browse the full model catalog at ollama.com/search. Tags like :8b or :14b pick the parameter size; most models default to a 4-bit quantization (Q4_K_M), which is the right starting point for local use. The quantization guide explains what that means for quality and memory.

Common Mistakes

• Downloading a model too big for your VRAM. If a model does not fit in GPU memory, Ollama offloads the rest to system RAM and generation slows to a crawl. Match the model to your hardware first — ollama ps shows whether a model loaded on GPU or CPU.
• Expecting the 671B “real” DeepSeek or a 70B model to run on a laptop. Those need data-center or multi-GPU memory. Stick to 7B–14B locally unless you have 24GB+ of VRAM.
• Forgetting the background service on Linux. If ollama run errors with a connection refused, the service is not running — start it with ollama serve or the systemd commands above.
• Confusing RAM with VRAM. 32GB of system RAM does not let you run a 32B model quickly; what matters for speed is GPU VRAM. See the VRAM requirements guide.
• Running the 8GB variant of a GPU and wondering why models stutter. 8GB is the floor for 7B models with no context headroom. 16GB is the comfortable sweet spot.

What To Do Next

• Pick a model for your hardware tier with the VRAM requirements guide and the VRAM calculator.
• Run a reasoning model: how to run DeepSeek R1 locally.
• Add a friendlier chat interface — Open WebUI or LM Studio — once the command line works.
• If models feel slow, the bottleneck is almost always VRAM. The best GPU for local LLMs guide covers the upgrade path.

Frequently Asked Questions

Is Ollama free?

Yes. Ollama is free and open source. The only cost is the hardware you run it on and the electricity to power it — there are no API fees or subscriptions.

Do I need a GPU to use Ollama?

No. Ollama runs on CPU by default and 7B–8B models are usable that way on 8GB of RAM. A GPU mainly makes generation faster; it is recommended but not required.

Does Ollama use my GPU automatically?

Yes. On Apple Silicon, current NVIDIA drivers, and supported AMD ROCm setups, Ollama detects and uses the GPU at runtime with no manual configuration. Run ollama ps to confirm a model loaded on GPU rather than CPU.

Where does Ollama store models?

By default, models live in your home directory (for example ~/.ollama/models on macOS and Linux). They can be several GB each, so use ollama list and ollama rm to manage disk space.

Can I run Ollama offline?

Yes. After a model is downloaded, Ollama runs entirely offline — nothing leaves your machine. You only need internet for the initial install and model pulls.

---

Last updated: May 2026. Install commands reference the official Ollama documentation at docs.ollama.com. Model sizes and tags reflect the official Ollama library at time of publication.