Ollama installs in one command and runs Llama 3, Mistral, Phi-3, Gemma, and 100+ other models locally. No API key. No cloud. No monthly bill. Your data never leaves your machine.
Type a prompt in your terminal or hit the local REST API from your own code. Llama 3 answers. Nothing goes to OpenAI, Google, or anywhere outside your machine.
Ollama is a small binary that manages model downloads and runs an inference server in the background. It takes one command to install on every OS.
Or download the .dmg from ollama.com if you prefer a GUI installer. The DMG version adds a menu-bar icon.
Now grab a model. Llama 3.2 3Bis the right starting point — it's fast, accurate enough for most tasks, and downloads in under 2 minutes on a decent connection.
Other popular picks — run any of these instead if you prefer:
Now talk to it. Run the model by name — Ollama starts an interactive session. Type a prompt, press Enter. Type /bye to exit.
Useful one-liner patterns (pipe text into the model):
Ollama runs a local REST API at http://localhost:11434. It's OpenAI-compatible, so any code that calls OpenAI can call Ollama with one line changed. You can call it with curl, Python, Node, or anything else.
OpenAI-compatible endpoint (drop-in for most SDKs):
You now have a local AI server running on your machine. No API key. No cloud. No monthly bill. 100+ models a ollama pull away.
The 5-minute estimate assumes a fast internet connection for the model download and an M-series Mac or recent CPU. On an older Intel Mac, first-token latency for llama3.2 is about 3–4 seconds. On Windows with an NVIDIA GPU, it's under 1 second.
We tested Mac (M3), Windows (RTX 3080), and Ubuntu (A100). The 93% first-time success rate reflects the install — the main failure mode on Windows is the PATH not updating until you open a new terminal window after installing.
Model quality varies a lot. Llama 3.2 3B is good for summaries, Q&A, and simple code. For complex reasoning or long-form writing, pull the 8B or 70B variant. The 70B needs 48GB RAM — most laptops top out at 16GB or 32GB.
ollama serve in one terminal, then ollama run llama3.2 in another. On Mac/Windows the installer usually starts it automatically on login.ollama rm modelname to delete models you no longer need. Run ollama listto see what's installed.ollama run phi3instead — it's 2.3GB and designed to run fast on CPU. The 70B models need 48GB RAM to fit in memory; below that they page to disk and crawl.Tomorrow's best AI project. Same kind of guide. Free. No spam ever.