Skip to content

Ollama

Pulls quantised models and serves them over a local HTTP API, with a menu bar app and a CLI.

Homebrew cask Homebrew formula Direct download

What it does

Ollama is a model runner: ollama pull fetches a quantised model, ollama run gives you a chat prompt, and a local server on 127.0.0.1:11434 exposes an HTTP API that other tools point at. Models are described by a Modelfile — base model, system prompt, parameters — which makes a customised variant a three-line text file.

ollama pull llama3.2
ollama run qwen2.5-coder:7b
ollama list
curl http://localhost:11434/api/generate -d '{"model":"llama3.2","prompt":"hi"}'

On Apple Silicon it uses Metal, and unified memory means the usable model size is roughly “RAM minus what you are using” — a 32 GB Mac comfortably runs 7B–14B models quantised.

Two installs, one thing

The cask (ollama-app) installs the menu bar app with a chat window and a bundled server. The formula installs just the CLI and server binary, which is what you want if you run it under launchd or only ever call the API. Installing both leaves two copies of the binary, so pick one.

Notes

The API is OpenAI-compatible at /v1, so most clients written against OpenAI work by changing the base URL. Models land in ~/.ollama/models and are large — check that before wondering where the disk went.

Alternative to

AlternativeTypeTrade-off
LM StudioFreeA graphical model browser and chat client over the same GGUF models
llama.cppOpen sourceThe inference engine Ollama wraps; more control, more flags
Jan and GPT4AllFreeDesktop apps aimed at running local models with no terminal
vLLMOpen sourceThe serious serving option when throughput and batching matter more than convenience

Install

brew install --cask ollama-app   # menu bar app + server
brew install ollama               # CLI and server only

Links

This post is licensed under CC BY 4.0 by the author.
Last updated on