Free guide
Can you run your own AI?
AI that runs on your own computer — at the office or at home — with your documents never leaving. Here is what it takes, what it does well, and when it is not worth it. Tick what you want it to do — at work and, if you like, at home too — and in a couple of minutes you will know, task by task, what you can do and on which machine.
The check
What it is, and what it isn’t
Local AI is a model you download and run on your computer, like any other program. When you ask it something, nothing leaves for the internet: not your question, not your documents.
It is not a stripped-down version of anything, nor a toy. It is the same class of technology, in a size that fits your machine.
What it is not: entirely free. The software costs nothing, but the machine and the time to set it up and teach it do.
What it takes, in two rules
On a Mac with Apple silicon, the computer’s memory is what counts, because it is shared with the graphics. It is the machine that gives the best value for this.
On a PC, the graphics card’s memory is what counts. The computer’s barely matters. Without a dedicated card it still works, on the processor, but slowly.
And the rule that covers both: the model has to fit in about two thirds of that memory.
| Memory available | Model size | What it reaches |
|---|---|---|
| 8 GB | Small | Simple, tightly defined jobs |
| 16 GB | Medium | Reading documents, pulling out data, answering about them |
| 32 GB | Large | Also writing and reasoning fluently |
| 64 GB or more | Genuinely big | More than most people need |
With names, as of September 2026. The model a local AI tool suggests by default for each machine:
| Machine | Default model |
|---|---|
| Mac 8 GB | Qwen3.5 4B (2,6 GB) — very tight |
| Mac 16 GB | Ornith-1.5 9B (5,5 GB) |
| Mac 24 GB | Ornith-1.5 9B (5,5 GB) |
| Mac 32 GB | Qwen3.8 27B (15,7 GB) — the most capable in the catalogue |
| PC 8 GB | Qwen3.5 2B (1,2 GB) — token only |
| PC 16 GB | Qwen3.5 4B (2,6 GB) |
| PC 32 GB | Ornith-1.5 9B (5,5 GB) |
| PC 64 GB | Qwen3.8 27B (15,7 GB) |
Names change every few months and this table will age; what lasts is the size rule. On a PC with a graphics card, the card’s memory is what counts.
What it is done with
The two tools almost everyone uses, both free:
- Ollama — driven from the command line, and the one used when it has to connect to other software.
- LM Studio — has a window and buttons. If you want to try it yourself this afternoon, start here.
Installing them takes minutes and your technician can do it perfectly well. That is not where the difficulty lies.
What works well and what doesn’t
Works very well: reading documents and pulling out specific data, answering questions about your own paperwork, classifying, sorting, summarising and preparing repetitive drafts.
Works so-so: writing text that goes out with your name on it. It writes correct, flat prose; it needs someone to go over it.
Works badly: asking it to know a regulation and judge by it. This is where people get disappointed and give up.
And here is the trick that changes everything: you don’t ask the model for the judgement, you give it in writing. Instead of "is this certificate correct?", you ask it to extract the data and compare it against a checklist written by someone who knows. With that, a medium model does work a huge model would not do without the list.
And at home? It can run the smart-home too
If you already have a smart-home system in place — lights, climate, blinds, alarm — a local model can be the thing that understands you when you speak in your own words: "close the south blinds, the sun is on them".
The interesting part is not the convenience, which any commercial speaker already gives you. It is that what happens in your home does not leave your home: not who you talk to, not what time you get in, not when nobody is there. And it runs on a modest computer, because these are short commands.
Two honest warnings: the smart-home has to exist already — this does not install it — and it is different work from the document side, because it connects to another system and is tested separately. If you are interested, say so when you write and we will look at it for what it is: a separate commission.
When it is not worth it
If your data is not sensitive and the volume is small. The cloud will be cheaper and work better. Setting it up in-house is paying for peace of mind you already have.
If nobody is going to maintain it. It has to be installed and updated now and then. With nobody on it, it ends up switched off in a corner.
If what you want is something to decide for you. AI prepares and flags. Whoever signs, signs.
And if you want us to do it
Five steps, and the second is the one that really counts:
- Choose the job. Which repetitive work is worth it and which isn’t.
- Write the criteria. Get out of the head of whoever does it today the list of what must be checked. An IT person cannot do this.
- Configure it. On your machine or in the cloud, depending on the data, if you hire us. Your technician installs the equipment.
- Test it on real cases until it gets them right, and write down where it fails.
- Teach the people and make clear what a person reviews and signs.
We don’t sell hardware: we neither buy nor supply computers or devices. We tell you whether what you have will do and, if not, what to ask your technician for.
And if you got this far
Does what you publish with AI need labelling? Since 2 August 2026, some of it does. If you work with data that cannot leave, this is the other half of compliance.
And can AI assistants find you? A different question in a different place, but it takes 30 seconds to check, free.
Frequently asked questions
Is local AI worse than ChatGPT?
For open conversation, yes: the big cloud models are in another league. But for reading your documents and pulling data out of them, a well-prepared medium model does the job just as well, and your paperwork never leaves the building.
Do I need to buy a new computer?
Often not. A recent machine with 16 GB already handles document work, which is most of what gets commissioned. First work out what you want to do; the machine is decided after that, not before.
Can it read scanned PDFs?
Yes, but with a character-recognition pass first, and the result is never perfect with stamps, tables and signatures. It is the first thing to check, because it changes the whole job.
Can it be taught our documents?
Yes, and that is exactly what should be done: you give it the documents, the procedures and the house criteria so it consults them when answering. That is not retraining the model — that takes weeks and is almost never needed — it is explaining who you are and how you work.
Can it validate a certificate or an official document?
It can prepare one: extract the data, compare it against a checklist and flag what does not add up. A person signs the validation. A system that said "this is correct" with nobody reviewing it would be a problem, not a solution.
How much does it cost?
The software costs nothing. What costs is deciding which job is worth it, writing the criteria and teaching the people. That depends on the job, which is why there is no price on this page: each case gets looked at.
Last reviewed: 2026-09-27 · This page carries model names, requirements and prices, and all of that ages: that is why we review it monthly. What does not expire is the size rule.