Why this choice? Because large data centres clear forests, take up land, use a great deal of water, and demand ever more electricity. Choosing a small, local AI means refusing that race.
The rule of sobriety. Not the most powerful. Not the best known. The smallest one that does the job — and you stop there, for the planet.
GPT4All
To get started. Guided graphical interface, no command line. Free and open source (MIT licence).
Go to the official site and download the version for your system.
Windows, macOS or Ubuntu. The site usually detects the right link automatically.
Install the software and open it.
Add a model from the built-in gallery.
Choose a model marked “download”. Avoid those that ask for a key or an online account: they go through the internet and are not local.
Some models in the gallery do not run locally. Read the label before downloading.
Start a conversation and ask a simple question.
For example: “Explain in ten lines what a local AI is, to a parent.”
Check that the computing is really local.
Turn off Wi-Fi, then ask the question again. If the AI answers, the computing happens on your computer.
This test proves the computing is local. It does not guarantee your data will never be sent. Turn off cloud options in the software settings.
LM Studio
Polished interface, a larger model catalogue. Free for personal and professional use.
Go to the official site and install LM Studio.
Look for a small model in the catalogue.
Guide by your memory (RAM) — 8 GB: 1B to 3B models · 16 GB: 3B to 7B · 32 GB: up to 13B.
Download the model, then load it in the software.
Open the chat tab and ask a question.
Check that the computing is really local.
Turn off Wi-Fi and check that the AI still answers.
Turn off server mode or online options in the settings if you see them.
Ollama
More technical: you type commands in the Terminal. Free and open source (MIT licence). On Apple Silicon Macs, Ollama uses MLX under the hood.
Go to the official site and install Ollama.
Open a terminal and check the installation.
ollamacopy
Run a small model. The download starts automatically.
ollama run phi3copy
Ask your question directly in the Terminal. To quit:
/byecopy
Check that the computing is really local.
Turn off Wi-Fi and restart the model you already downloaded.
The model stays cached on your computer. It no longer needs the internet to work.
Useful commands
ollama listcopy
ollama pull mistralcopy
ollama run mistralcopy
MLX
Apple's native engine for its M chips. Maximum performance on Mac. Requires Python and the Terminal. Does not work on older Intel Macs.
Check that you have an Apple Silicon Mac.
Apple menu → About This Mac. The chip must be M1, M2, M3 or M4.
Open the Terminal and create an isolated workspace.
python3 -m venv mlx-env source mlx-env/bin/activatecopy
You will see (mlx-env) appear at the start of the command line.
Install MLX.
pip install -U mlx-lmcopy
Run a first model. The download starts automatically.
mlx_lm.generate --model mlx-community/Llama-3.2-3B-Instruct-4bit --prompt "Hello, who are you?"copy
Next time, the model is already cached on your Mac.
For a continuous conversation:
mlx_lm.chat --model mlx-community/Llama-3.2-3B-Instruct-4bitcopy
8 GB of RAM → a 3B model is comfortable. 16 GB → up to 13B. No data leaves your Mac.
This workshop extends the site's thinking on digital autonomy and sobriety. Read: Responsible digital autonomy →
Tools and models change fast. Apologies in advance if some necessary updates are missing here. June 2026 version.