Skip to content

Instantly share code, notes, and snippets.

@drbenvincent
Last active May 18, 2026 09:36
Show Gist options
  • Select an option

  • Save drbenvincent/894c0b3e83cf9808fb9c052ccd5270bb to your computer and use it in GitHub Desktop.

Select an option

Save drbenvincent/894c0b3e83cf9808fb9c052ccd5270bb to your computer and use it in GitHub Desktop.
Local LLM install instructions - PyMC Labs retreat 2026

Before The Retreat

If you want to play with local LLMs on the retreat, please install Ollama and download at least one model before you travel. Pulls are still large enough that, if everyone downloads on the venue Wi-Fi in Barcelona, the network will die.

I only have capacity to test this on macOS. If you are using Linux or Windows, please follow the official Ollama installation instructions at https://ollama.com/download.

Install Ollama On macOS

Install Ollama from the terminal:

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

This installs the ollama command-line tool and the Ollama desktop app.

Check which models are already available locally:

ollama list

The list will usually be empty the first time you run this.

Download The Workshop Model

The demo reads llm.ollama.model from config.yaml. Use the same tag in both places.

Default (recommended): gemma4:e4b — tested on an M4 MacBook Air with 16GB RAM. In Ollama this is an ~8B model at Q4 quantization; it is a comfortable fit on 16GB Apple Silicon.

# Workshop default: tested on 16GB Apple Silicon; good quality for summarisation and JSON
ollama pull gemma4:e4b

If you have 8GB RAM or an older laptop: pick one smaller model instead. Summaries and JSON may be a little less reliable, but the demo still runs.

# ~2GB download; strong instruction-following for its size (helps with structured JSON output)
ollama pull qwen2.5:3b
# ~2GB download; widely used 3B baseline if the Qwen tag is slow or unavailable on your machine
ollama pull llama3.2:3b

If you want to experiment with other models later, a good approach is to give an LLM your hardware specs and ask for a shortlist of Ollama models that fit your machine.

Ollama stores downloaded models under ~/.ollama/models/.

FYI: You can kill ollama with pkill ollama, and you can remove models with ollama rm gemma4:e4b

Test it works

Just in case, you might want to close memory hungry apps first. The command below should let you confirm it works:

ollama run gemma4:e4b "What is your favourite colour?"

If you want to do some extra-curricular playing around...

Then downloading the Pi agent harness would be useful, see pi.dev

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment