Getting Started with Ollama

Installing Ollama

Ollama runs on macOS, Windows, and Linux. The installation is a single step on every platform.

macOS

Download the installer from ollama.com and open the .dmg file. This installs Ollama as a macOS application that runs in your menu bar and starts automatically on login.

Important: Do not install Ollama via Homebrew on macOS. The Homebrew package may install an x86 binary that runs under Rosetta 2 emulation — this binary cannot use Apple Metal GPU acceleration and will run 10–100x slower than it should. The .app installer from ollama.com ships as a universal binary, runs its server as an ARM-native process, and includes Metal GPU support out of the box.

Linux

Run the single-line installer in your terminal:

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

This works across most distributions (Ubuntu, Debian, Fedora, Arch). The script detects your hardware and installs the appropriate binary. NVIDIA and AMD GPU support is configured automatically if drivers are present.

Windows

Download the Windows installer from ollama.com. It installs Ollama as a background service and adds the ollama command to your terminal. Supports NVIDIA GPUs natively; AMD support on Windows is improving but NVIDIA is more reliable.

Your First Model

Open a terminal and run:

ollama run llama3.3

Ollama will download Llama 3.3 8B (around 4.7 GB at the default Q4_K_M quantization) and drop you into an interactive chat session. Type a message and press Enter to get a response. Type /bye or press Ctrl+D to exit the chat.

The first run takes a few minutes for the download. Every run after that starts in seconds.

Core CLI Commands

These are the commands you will use most:

• ollama pull <model> — Download a model without running it. Use to pre-load before you need it.
• ollama run <model> — Start an interactive chat with a model. Downloads it first if not already on disk.
• ollama list (or ollama ls) — Show all models downloaded on your machine.
• ollama rm <model> — Delete a model from disk.
• ollama ps — Show which models are currently loaded in memory.
• ollama stop <model> — Unload a model from memory (frees VRAM without deleting the file).
• ollama serve — Start the Ollama API server manually. On macOS the server starts automatically; on Linux and Windows you may need this.

Verifying GPU Acceleration

GPU acceleration delivers a 10–20x speed improvement over CPU inference. It is worth confirming it is working before assuming your performance is normal.

The easiest check: run ollama ps while a model is loaded. The output shows the model name and which hardware it is using. If you see your GPU name, acceleration is working.

You can also watch inference speed. With GPU acceleration, a 7–8B model should produce 20–50 tokens per second. If you are seeing 2–5 tokens per second, the model is likely running on CPU.

Platform-Specific Gotchas

macOS / Apple Metal: If you installed via Homebrew and inference feels slow, uninstall and reinstall using the .app installer from ollama.com. The Rosetta 2 binary does not use Metal. After reinstalling, Metal acceleration should work automatically on any M-series Mac.

AMD (Linux): Ollama maintains a list of officially supported AMD GPUs. Cards not on the list — including some RDNA 2/3 series cards like the RX 6600 and RX 6700 — will silently fall back to CPU. You can override this with an environment variable before starting Ollama:

HSA_OVERRIDE_GFX_VERSION=10.3.0 ollama serve

Use 10.3.0 for RDNA 2 cards (RX 6600/6700/6800 series) and 11.0.0 for RDNA 3 APUs. This is an unofficial override — test your results with ollama ps to confirm GPU is being used.

NVIDIA: Ensure your NVIDIA drivers are up to date. Ollama detects CUDA automatically when drivers are correctly installed. If the GPU is not detected, the most common fix is updating to the latest driver from nvidia.com and restarting the Ollama service.