Get Started
← Own your AI · Mission 1 of 31

Run a ChatGPT-style AI on your own laptop

Google's Gemma 4, running on your machine with Ollama. No account, no subscription, and it keeps answering with the Wi-Fi switched off, because nothing you type leaves your computer.

15 minBeginnerTools: FreeFree mission
Tested 7 Oct 2026 · Ollama 0.40, Gemma 4 e2b, 16GB laptop with no graphics card
Ollama installedthe engine that runs AI models on your own machine
Gemma 4 downloaded and answering you in the terminal, even with the internet off
Real speedabout 26 tokens a second on a normal laptop with no graphics card
Before you start
  • 8GB of RAM minimum, 16GB recommended
  • About 7GB of free disk space
  • Level 0 isn't required for this one. All you need is a terminal window.
01/05Install Ollama0 done
STEP 01 of 05~5 min

Install Ollama

Ollama is the free, open-source engine that downloads and runs AI models for you.

  1. Download the installer from the official site, ollama.com/download ↗, and run it. Or install it with one command:

    Windows
    winget install --id Ollama.Ollama -e
    Mac (or use the download)
    brew install ollama
    Linux
    curl -fsSL https://ollama.com/install.sh | sh

  2. The Windows installer is about 1.6GB, so give it a few minutes. When it's done, Ollama runs quietly in the background: on Windows you'll see its llama icon by the clock, on a Mac it sits in the menu bar, and on Linux the install script sets it up as a background service.

  3. Open a new terminal and check it's there:

    Terminal
    ollama --version
    Output
    ollama version is 0.40.3
✓
Checkpointollama --version prints 0.20.7 or newer (0.30.5 or newer if you plan to use gemma4:12b).
If it breaksIt says the model needs a newer version of Ollama + 2 more
It says the model needs a newer version of Ollama
Gemma 4 needs Ollama 0.20 or newer, and 0.20.7 fixed answer quality for the small e2b and e4b models with thinking off. gemma4:12b arrived in 0.30.3 and needed 0.30.5 to stop crashing on Windows and Linux. Run the installer or winget upgrade --id Ollama.Ollama -e again (brew upgrade ollama on a Mac, the same curl line on Linux), then check ollama --version.
'ollama' is not recognised
Close every terminal and open a new one. If it still fails, restart the PC. The installer adds Ollama to your PATH, but only new windows pick that up.
'could not connect to a running Ollama instance'
The background app isn't running. Open Ollama from the Start menu (or Applications on a Mac) and try again. On Linux, run ollama serve in a separate terminal window and leave it open.
Still stuck? Ask the community
STEP 02 of 05~2 min

Pick a model that fits your machine

Bigger models are smarter but need more memory. Pick the wrong one and it crawls.

A model is the AI's "brain" file. Ollama has hundreds. For a normal laptop, start with Google's Gemma 4. The small versions were designed to run on laptops and phones.

  • 8 to 16GB RAM, no graphics card: gemma4:e2b (4.6GB download). Start here.
  • 16GB+ RAM: gemma4:e4b (6.6GB). Smarter, a little slower.
  • A gaming PC with 12GB+ of graphics memory: gemma4:12b (7.7GB).
TipAlways type the size. ollama run gemma4 with no size pulls the latest tag, which points at e4b (6.6GB), not the smaller e2b. Every command in this mission names the size, so you get the same model the lesson was written with.
TipNot sure how much RAM you have? On Windows: Settings → System → About → Installed RAM. On Mac: Apple menu → About This Mac.
✓
CheckpointYou've chosen your model. If in doubt, use gemma4:e2b.
Stuck? Ask the community
STEP 03 of 05~6 min

Download it and say hello

One command downloads the model and opens a chat.

  1. Terminal
    ollama run gemma4:e2b

  2. The first time, it downloads the model (about 4.6GB, roughly 5 to 15 minutes on home broadband). After that it starts in seconds. When you see >>>, type a message and press Enter:

    Output
    >>> Write a 3-line text to my landlord saying the boiler is broken.
Thinking...
...
...done thinking.

Here are a few options, depending on how urgent and formal you want to be:

**Option 1: Clear and Direct (Recommended)**

> Hi [Landlord's Name],
>
> I am writing to let you know that the boiler is currently broken and not working.
>
> Could you please arrange for a repair as soon as possible?
...

  3. Now the good bit. Turn your Wi-Fi off (or pull the cable), type another question and press Enter. It still answers, because the whole model is on your disk and running on your own processor.

  4. Type /bye to leave the chat. Run the same command again any time to come back.

✓
CheckpointThe model answers you in the terminal, and keeps answering with the internet off.
If it breaksThe download stalls or fails + 1 more
The download stalls or fails
Run the same ollama run command again. It picks up where it left off.
It's painfully slow, or your PC freezes
The model is too big for your RAM. Press Ctrl + C, then remove it with ollama rm and the model name, and try gemma4:e2b. Closing Chrome tabs frees up a surprising amount of memory.
Still stuck? Ask the community
STEP 04 of 05~2 min

Speed it up: skip the thinking

Gemma 4 thinks before it answers. Great for hard questions, slow for simple ones.

  1. By default Gemma 4 "thinks" first: you'll see Thinking... before the answer. On a 16GB laptop with no graphics card, the boiler text above took 18 seconds with thinking on and under 5 with it off, because the thinking alone ran to more than 300 words. For quick jobs, turn it off:

    Terminal
    ollama run gemma4:e2b --think=false

  2. Add --verbose to see exactly how fast it runs on your machine. Here's a real run on a 16GB Intel laptop with no graphics card:

    Output
    total duration:       4.716286s
load duration:        17.0933ms
prompt eval count:    25 token(s)
prompt eval rate:     43.48 tokens/s
eval count:           134 token(s)
eval duration:        4.576452s
eval rate:            29.28 tokens/s

  3. About 29 tokens a second is roughly 20 words a second, faster than you can read. A short answer took under 5 seconds. The very first run after a restart is slower: loading the model from disk took 49 seconds here, then every answer after that started straight away.

    See what's loaded
    ollama ps
    Output
    NAME          ID              SIZE      PROCESSOR    CONTEXT    RUNNER      UNTIL
gemma4:e2b    b37049369adf    970 MB    100% CPU     4096       llamacpp    4 minutes from now
    PowerShell running ollama run gemma4:e2b with thinking off and --verbose: three short landlord messages about a broken boiler, then timing lines ending in eval rate 29.28 tokens/s, then ollama ps showing gemma4:e2b at 970 MB on 100% CPU
    Gemma 4 e2b on an Intel Core Ultra 7 laptop with 16GB and no graphics card: 134 tokens in 4.7 seconds.
✓
CheckpointYou've had an answer with thinking switched off, and you know your own speed.
Stuck? Ask the community
STEP 05 of 05~1 min

Manage your models

Three commands keep your disk tidy and your models up to date.

This is what makes it different from ChatGPT: you can paste private things like payslips, contracts or health questions, and they never go to a server. To look after it:

  • ollama list shows the models you've downloaded
  • ollama rm gemma4:e2b deletes one to free up space
  • ollama pull gemma4:e2b updates it to the latest version
✓
Checkpointollama list shows gemma4:e2b.
Stuck? Ask the community

Your first wins

It works. Now do something with it.

Summarise something private

Paste a contract, a tenancy agreement or a payslip into the chat. It never leaves your laptop.

Prompt
Summarise this in plain English. Then list anything unusual, anything I should push back on, and any dates or deadlines I need to know about:

[paste the text here]

Use it as a writing partner

Small local models are great at rewriting. They're weaker at facts, so check anything important.

Prompt
Rewrite this message so it's shorter, friendlier and more confident. Give me 3 versions:

[paste your message]

Now give it a proper app

The terminal works, but a chat window is nicer. Build your own private ChatGPT puts a ChatGPT-style interface on top of Ollama: chat history, file uploads, and switching between models.

Undo it cleanly

How to remove everything this mission installed.

Delete the model first (this frees the disk space), then uninstall Ollama:

Windows
ollama rm gemma4:e2b
winget uninstall --id Ollama.Ollama -e

On a Mac, quit Ollama from the menu bar and drag it from Applications to the Bin. Downloaded models live in a hidden .ollama folder in your home folder. Delete it to remove them all.

On Linux, remove the model with ollama rm gemma4:e2b, then follow the uninstall steps in Ollama's Linux docs ↗: they stop the service and remove the program and its models.

Next mission · Own your AIThe tiny AI model that got basic maths wrong, and the one to use instead
→