AI at home, explained without jargon

First, the map: what leaves home and what does not

Not all AI works the same way. The important difference is not the brand, it is where each thing is processed: on another company’s server, or inside your own computer.

Tool	Where it is processed	Your data
ChatGPT (web or app)	OpenAI’s servers, in the United States	Leaves home. May be used to improve models depending on account settings
Copilot (chat)	Microsoft’s cloud	Questions and attachments leave home
Copilot+ PC: local features (e.g. Recall)	The neural chip in your own PC	Stay on the disk, encrypted; Microsoft says it cannot read them
Local AI (LM Studio, Ollama, Jan…)	Your Mac or PC	Never leaves home. You do not even need internet once it is installed

The honest nuance: new Windows machines sold as “Copilot+ PCs” do run AI inside the computer for certain features. But the Copilot chat, the one people use to ask things, still sends your question to Microsoft’s cloud. “Copilot” does not mean “local”.

How do I know if my machine is up to it?

The quick rule, without opening any diagnostic program:

• Mac with an Apple chip (M1, M2, M3, M4…) and 16 GB of memory: yes, comfortably, for chatting with open models. With 8 GB it also works, with smaller models.
• Windows PC with 16 GB of RAM: yes, fine for small and medium models. If it also has an NVIDIA graphics card, much faster.
• A laptop more than 6 or 7 years old with 8 GB of RAM: you can try small models, but it will be slow.

On a Mac, memory is shared between the processor and the graphics: a 16 GB Mac can dedicate a good part of it to AI. That is why recent Macs are such good candidates.

Try it yourself, step by step

This is not a service offer: it is what anyone with a reasonably modern machine can do on a Saturday morning. Everything below is free.

If you have a Mac

1. Go to lmstudio.ai and download the macOS version (choose “Apple Silicon” if your Mac has an M chip).
2. Open the app. In the model browser (the telescope icon, “Discover”), type the name of a small, well-proven model: for example Qwen3 8B with 16 GB of RAM, or Phi-4 Mini with 8 GB.
3. Press Download and wait (it will take between 2 and 8 GB on disk, depending on the model).
4. Open the chat, select the model and ask whatever you like. Nothing leaves the Mac. You can disconnect the internet and it keeps working.

If you have a Windows PC

1. The closest route is downloading LM Studio for Windows, or Ollama (ollama.com), which works from a command window but is very simple.
2. With Ollama: install it, open its browser library, or type ollama run qwen3:8b in a terminal and it downloads and starts the model itself.
3. If the PC has an NVIDIA card, the app will detect it and run faster; if not, it works on the processor — slower, but it works.

What you can (and cannot) expect

• Yes: drafting and correcting texts, summarising documents, translating, answering questions, transcribing notes, coding with help, chatting with nobody reading over your shoulder.
• It depends: home models do not search the web by themselves (you have to enable it or connect a search tool), and they do not generate images at the quality of paid tools.
• No: a local model does not “remember” past conversations between sessions unless you add a memory layer, and it does not match the frontier cloud models on the hardest tasks. For day-to-day work the gap is small; for an extreme reasoning exam, it is still there.

And if I want to go further?

Once you have tried the chat, the natural next step is a search engine over your own documents: the AI answers from your files and cites the source, all inside your home. The tools exist and are free; setting it up well (fast answers, correct citations) takes hours. If you have delicate data — clients, patients, legal information — and you want the system properly built and auditable, that is where a professional has something to say. We do it for businesses and firms in Barcelona; for home use, this guide is enough to start.

Quick questions

Is it legal to run AI at home? Yes. No rule forbids running a program on your own computer. What the law regulates is what you do with other people’s data.

Are the models free? The most popular ones (Qwen, Llama, Gemma, Mistral, DeepSeek) are downloaded and used at no cost. The only costs are electricity and disk space.

And the quality, compared to ChatGPT? Today’s open models are at a very high level for normal tasks: drafting, summarising, translating, coding. They lose, frankly, on the most extreme reasoning tasks and on current events (they have learned nothing after their cut-off date).

If I already own a powerful machine, why would I pay you? You do not have to pay anything to use AI at home: that is what this guide is for. Asking for help makes sense when the system will be used by people in an office, when sensitive client data is involved, or when you want it to run on its own and in an auditable way.