Local large language models (LLMs) run on your own computer. They have changed a great deal since the first edition of this workshop in 2024. Privacy and cost are still good reasons to run a model locally. What has changed is that many capable models now run on an ordinary laptop and come with permissive licenses. A few are fully open. Their weights (the learned numbers inside the model) are published, and so are the training code and the training data or the scripts to rebuild it. For libraries, this means you can check provenance and reproducibility on your own machine. It also means you do not depend on a single vendor. This third edition is built around Apertus, a fully open multilingual model family from the Swiss AI Initiative (EPFL, ETH Zurich, CSCS).

Part 1 - Fundamentals

How LLMs work (parameters, tokens, quantization, context window). The difference between open source, open weights, and fully open models. Installing a local runtime (Ollama or LM Studio), the program that runs the model on your computer. Choosing a model size for your hardware and writing effective prompts. No programming experience is required for this part.

Part 2 - Working with your data

Working with LLMs from Python. Generating structured output (answers in a fixed format, usually JSON) and validating it. Embeddings (lists of numbers that represent the meaning of a text) and semantic search over controlled vocabularies. Retrieval-augmented generation (RAG): fetching relevant passages from your own documents and giving them to the model with the question. Basics of tool use, where the model asks your program to run a function, for example querying a SPARQL endpoint or an authority file. A person reviews the results (a human in the loop).

This is a hands-on workshop. You can follow it with the graphical tools only, run the notebooks (the Python code for Part 2), or work through the materials afterwards. The four hours are split into two halves. Part 1 is for beginners and needs no prior experience. Part 2 assumes you can already run a local model and builds on it with Python. Each half stands on its own: you can leave at the break, or join at the break for Part 2.

What you need

A computer running Windows, macOS or Linux with 20 GB of free storage. 8 GB of memory (RAM) is enough for the smaller model; 16 GB lets you run the larger one. You need an internet connection to download the models, unless you have downloaded them in advance. The steps are on Before you start. For Part 2 you also need Python 3.12 or 3.13. The Python package manager uv can install it for you.

When

Monday 16 November 2026, 04:00-08:00 UTC, with a ten-minute break at 06:00. Zoom opens at 03:30 UTC for installation help. Local times, the Zoom details and the setup steps are on Before you start.

Three ways to take part

Way	What you do	What you need
GUI only	Part 1 fully, Part 2 as a listener	Ollama or LM Studio and the model
Run the notebooks	Follow Part 2 hands-on	Python 3.12 or 3.13, uv, the notebooks repository, Ollama
Read afterwards	Work through the executed notebooks and these pages	Nothing to install

Before the workshop

During the workshop

Facilitator: Nishad Thalhath, RIKEN Center for Integrative Medical Sciences. Materials CC BY 4.0, code MIT.