Home / Learn / Data & AI / Offline AI / Lesson 1.4
All lessons

Module 1: Welcome, and why run AI yourself · Lesson 1.4 · about 15 min

Pick your track: laptop, desktop or team server

Answer four questions about your computer, your users, your budget and your goal, and get a suggested path through the course.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 1 · about 15 minutes · read first: 1.1 to 1.3. You’ll be able to name your track and the model size that suits your machine.

Your memory decides which models you can run, so start with the computer you have and pick the track that matches it.

1. Why this matters

The course covers everything from a phone to a team server. You do not need all of it at once. Picking a track tells you which lessons to take first. It also stops you buying hardware you do not need.

2. The three tracks

  • Laptop track. You use the computer you already have, usually 8 to 32 GB of memory. You run small and mid-sized models. Most people start here, and many stay here.
  • Desktop track. A desktop with a graphics card, a Mac with a lot of memory, or one of the new 128 GB “AI boxes”. You run bigger models, faster, and for longer.
  • Team server track. One strong machine on the office or home network that several people use at once. You learn serving, logins and monitoring.

3. Find yours

Answer the four questions. The suggestion updates as you change any answer.

Try it: pick your track
What will you run it on?
Who will use it?
How much could you spend on new hardware?
What matters most to you?

Change any answer and the advice changes. Prices are rough guides from November 2026; Module 5 has the details for each country. Nothing you pick here is saved.

The model names in the suggestion are examples from late 2026. New models come out every few weeks. What stays true is the size class: how many billions of parameters fit your memory. Module 7 shows how to pick from whatever is newest.

“The B in 7B means gigabytes, so a 16 GB laptop runs a 16B model.” B means billion parameters (the numbers inside the model), not gigabytes. How many gigabytes those take depends on how precisely each one is stored. At 16 bits a 7B model needs 14 GB; squeezed to 4 bits it needs about 3.5 GB. A 16 GB laptop is comfortable with about 12B at 4 bits, not 16B.

Maths and principle

Principle: a model fits when its size in memory, plus some room to work, is less than your free memory.

Equation: memory needed ≈ parameters (billions) × bits per weight ÷ 8 GB, plus about 1 to 3 GB of working room.

Analogy: packing a suitcase. Parameters are the clothes; bits per weight is how tightly you fold them (Module 8). Fold tighter and more fits, but things crease: quality drops a little.

Try it: a 12-billion-parameter model folded to 4 bits per weight? (12 × 4 ÷ 8 = 6 GB, plus working room, about 7 to 9 GB: fine on a 16 GB laptop.)

Watch the suitcase get packed for that 12B model on a 16 GB laptop.

The laptop. 16 GB of memory. The operating system and your open apps already use about 4 GB, leaving about 12 GB.

The weights. 12 billion numbers at 4 bits each. 4 bits is half a byte, so 12 × 4 ÷ 8 = 6 GB.

Working room. The model also needs space for the conversation it is reading (Lesson 3.3): about 2 GB for a normal chat.

Check the fit. 4 + 6 + 2 = 12 GB of 16. It fits, with 4 GB to spare. Laptop track, models up to about 12B.

Try a bigger one. A 27B model at 4 bits is 13.5 GB of weights alone. 4 + 13.5 + 2 = 19.5 GB: it does not fit. That size belongs to the desktop track.

P
parameters, in billions: the "7B" or "12B" in a model's name
b
bits per weight: how precisely each number is stored (16 for full size, 4 for a typical squeezed model)
P × b ÷ 8
the weights' size in gigabytes (8 bits make 1 byte)
working room
about 1 to 3 GB more for the conversation and the program itself
Physical meaning
The model has to sit in fast memory while it works, like a cook who needs every ingredient on the counter. If it does not fit, it either refuses to load or crawls along, borrowing slow storage.

4. Worked example

Made-up numbers. Aisha has a laptop with 16 GB of memory, works alone, has no budget for now, and cares most about writing code. The tool suggests the Laptop track, models up to about 12 billion parameters, and the short path through Modules 9 and 11, then lesson 13.3 on a local coding assistant. Three weeks later she finds the 12B model is good enough for most coding questions and keeps one cloud plan for the hard ones.

5. Practice

  1. Do the questionnaire for your own situation and write down your track.
  2. Using the box, how much memory does a 27-billion-parameter model need at 4 bits per weight?
  3. Change only the “Who will use it?” answer to a team. What changes in the advice, and why?
Answers
  1. Keep the note; lesson 18.12 (capstone) asks for it.
  2. 27 × 4 ÷ 8 = 13.5 GB, plus working room: about 15 to 17 GB. It needs a 24 GB machine or graphics card to run comfortably.
  3. It moves to the Team server track, because several people asking at once needs a machine set up for many users.

New words: track (your route through the course), size class (models grouped by billions of parameters), bits per weight (how precisely each number in the model is stored).

  • Three tracks: laptop (the machine you have), desktop (bigger and faster), team server (many users at once).
  • How much fast memory you have decides which models you can run.
  • Memory needed ≈ parameters (billions) × bits per weight ÷ 8 GB, plus 1 to 3 GB of working room.
  • A 16 GB laptop is comfortable with models up to about 12B at 4 bits per weight.
  • Start on what you have; it tells you what, if anything, is worth buying.

Check yourself

1What decides most strongly which models you can run?
2Five colleagues want to share one local model. Which track fits?
3You have a 16 GB laptop and no budget. What is a good first step?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.