Heidelberg AICurriculum
Track 19 · Beginner
19.13

The model on your bench

A 2 GB file that answers questions with your laptop and nothing else

2 lessons 2026-08-13 AI-generated

1Overview

A small open-weights model, downloaded once, running in a container on your own machine.

The bench ships with a small model that runs on the CPU of an ordinary laptop — no graphics card, no account, no internet. This chapter is about what such a model is genuinely good at, what it is not, and how to swap it for a different one. Knowing the difference is the difference between "local AI is useless" and using it for the half of the work it does well. → `local-private-models` is the wider landscape of models you can run yourself.

qwen2.5:3b is about 2 GB and answers on a CPU. That is a remarkable place to be: five years ago nothing this capable ran on a laptop at all. It is also genuinely limited, and the honest lesson is where the line falls. Good: summarising, extracting fields, rewriting, classifying, drafting, and every exercise in this course that is about mechanics rather than brilliance. Weak: long chains of reasoning, code of any size, anything where a subtle mistake is expensive.

1.2After this chapter you can
Say what a 3B model does well and where it fails
Swap the bench model for another one and feel the trade
Decide per task whether local is enough
1.3When to reach for it

Everything private, everything offline, and everything where a good-enough answer is genuinely good enough.

1.4Key parts

It answers through the gateway like any other provider — nothing about the apps knows it is local.

1.5Watch out

On 8 GB of RAM, run the model or a browser full of tabs, not both. Stop the model when you are using a cloud provider.

2Lessons 2

2.1 Find the edge of a small model, deliberately

A 3B model has read a lot and can hold little. Testing it on purpose is faster than discovering its limits inside a task you cared about.

Build your own sense of where local stops being enough.

Two models, one answering. The tile holds every model you have pulled and marks the one the gateway currently serves.
  1. Give it a summary task with a page of text. It will do well.
  2. Give it a extraction task — "list every date and what happened on it". Also well.
  3. Give it a three-step reasoning task — a puzzle, a proof, a refactor. Watch it produce something confident and wrong.
  4. Now switch to a big model and run the same three. The first two look identical. That is the whole lesson.
  • You'll see Two of three tasks where the free local model is indistinguishable from the paid one.
  • Takeaway Pick the model per task, not per person. Most course work runs fine on the small one.

2.2 Swap in a different model

The model app on the bench is Ollama. It can hold several models and answer with whichever you name.

Do this first Find the edge of a small model, deliberately

Pull a second model and compare the two on your own task.

Trydocker exec ollama ollama pull llama3.2:1b

in a terminal. 1b is smaller and faster than the shipped 3b — a good contrast.

  1. Pull it. The tile lists every model it holds, with the one currently answering marked.
  2. Point the bench at it through "Which AI answers" — the bench model option, with the new name.
  3. Run your own task on both. Speed against quality, measured on the thing you actually do.
  4. Disk matters. Each model is gigabytes; remove what you do not use with docker exec ollama ollama rm .
  • You'll see Both models on the tile, and a visible speed difference between them.
  • Takeaway Smaller is faster and dumber, and where the line falls depends on your task — so test it on your task.

💬 Discuss this chapter

Ask, share, or report — over on the Heidelberg AI community forum.