Explore AI

Who runs the model?

Almost every question about running AI yourself comes down to one decision, and most guides never state it plainly.

Whose computer is running the model?

Everything else follows from that — cost, privacy, quality, speed, and what you can actually build.

The two answers

Hosted. You type something, it goes to someone else's server, an answer comes back. Claude, ChatGPT, Gemini. You need a browser and nothing else.

Local. The model runs on the machine in front of you. Nothing you type leaves it. You need hardware, patience, and a willingness to troubleshoot.

What each one actually costs you

HostedLocal
Upfront costnone€0–2000+, or a machine you already own
Running costper token, or a monthly subscriptionelectricity, and your time
Your dataleaves your machinenever leaves it
Capabilitythe best available model, always currentcapped by your hardware
Speedfast, and someone else worries about itdepends entirely on your machine
Works offlinenoyes
Controlprompts onlythe weights, the context, everything
Setup effortnonereal, and occasionally maddening
Longevitythe provider can change or withdraw itwhat you downloaded keeps working

The honest part

Most comparisons of these two options are written by someone selling one of them — either a cloud provider, or a hardware-enthusiast site funded by affiliate links. This is the version that is not selling anything.

If your work is occasional, hard, or quality-critical, use a hosted model. The best hosted models are better than anything you can run at home, they cost nothing to start, and your time is worth more than the electricity saved. Nobody should feel bad about this answer.

If your work involves material you would not paste into a website, run it locally. Client contracts, medical records, a diary, source code under NDA, anything belonging to someone who did not consent. This is the strongest argument for local, and it is not really about AI at all — it is about where the data goes.

If your work is bulk and boring, run it locally. Transcribing four hundred recordings, classifying ten thousand support tickets, extracting fields from invoices. At volume, per-token pricing stops being convenient and starts being a line item. A local model on hardware you already own has no marginal cost.

If you want to understand the machinery, run it locally. You will learn more from a disappointing 4B model on a laptop than from a year of reading about frontier models.

Most people should do both

The mistake is treating this as a side. Sensible practice looks like:

  • Hard problems and writing you care about → a hosted model
  • Anything private → local, without exception
  • Anything large and repetitive → local, overnight, on a machine that can be left alone
  • Learning → local, and cheap, so that mistakes cost nothing

The rest of these appendices exist to answer the follow-up question: if I go local, what can my machine actually run, and what is it actually good at?

Where to go next