Ollama is a free, open-source program that lets you download an AI language model and run it directly on your own computer, instead of sending every question to a company’s servers. Install it, pull a model such as Llama or Mistral, and within minutes you have a private chatbot running offline — no subscription, no per-message fee, and no data leaving your machine.

What you need

Ollama runs on macOS, Windows, and Linux. The main requirement isn’t a powerful processor — it’s memory. A general rule of thumb: a model needs roughly as much RAM (or graphics memory) as its file size on disk. A smaller model in the 7–8 billion parameter range fits comfortably on a laptop with 8–16GB of RAM; larger, more capable models need proportionally more. A dedicated graphics card speeds up responses substantially and is detected automatically, but it isn’t required — Ollama falls back to running on the CPU alone.

Installing and running your first model

Getting started takes three steps:

  1. Install it. On macOS or Linux, paste the one-line install command from the official download page into a terminal; Windows users run a standard installer.
  2. Pull a model. Typing ollama pull llama3 (or the name of any model in Ollama’s library) downloads that model’s weights — typically several gigabytes — to your machine.
  3. Run it. Typing ollama run llama3 opens a chat prompt straight in the terminal. The same model is also exposed as a local web API, so developers can point their own applications at it instead of typing commands by hand.

How it works

Under the hood, Ollama is a friendly wrapper around llama.cpp, an open-source library that performs the actual model inference and works efficiently on ordinary consumer hardware, not just data-center chips. When you run a command, Ollama’s CLI sends the request to a small local server it starts automatically, which loads the model and generates a response. The company’s founders, Jeffrey Morgan and Michael Chiang, previously built Kitematic, a tool that made Docker easier to use before Docker acquired it — and Ollama applies the same idea to AI models: package them so that downloading and running one is as simple as pull and run, the same verbs Docker uses for software containers.

Tips and pitfalls

Match the model to your hardware — a model too large for your available memory will run painfully slowly or fail to load at all; smaller variants of the same model (labeled with lower parameter counts) trade some capability for speed. If you want access to larger, more capable models than your computer can run locally, Ollama also offers a paid cloud tier: as of July 2026, per Ollama’s pricing page, the free plan covers local use and light cloud usage, while Pro ($20/month) and Max ($100/month) plans unlock running several large cloud models at once. One real security pitfall: Ollama’s local server is meant to stay on your machine, but researchers found roughly 175,000 Ollama servers exposed to the public internet in early 2026 because their owners had configured them to accept connections from any address. Leave the default local-only setting in place unless you specifically know how to secure a public endpoint.

In the news

Ollama’s approach has caught on fast: the company recently raised $65 million and said it has reached 8.9 million developers, underscoring how much demand there is for running AI models outside the big providers’ clouds.

FAQ

Is Ollama really free? Running models locally is free and always has been — you’re just using your own computer’s resources. Ollama’s paid tiers only apply if you want to run larger models on its cloud infrastructure.

Which models can I run? Ollama’s library includes open-weight models such as Llama, Mistral, Gemma, Qwen, and DeepSeek, each available in several sizes.

Do I need a GPU? No. A GPU makes responses faster, but Ollama runs on the CPU alone if no compatible graphics card is present.

Is my data private? Yes, when run locally — prompts and responses never leave your computer unless you deliberately use Ollama’s cloud models.

Sources: Ollama’s official site, download page, and pricing page; the Wikipedia entry on Ollama.