Running a language model on your own machine has moved from an experiment to something a non-specialist can do in half an hour.
What you need
Memory is the binding constraint. A model in the seven to fourteen billion parameter range needs roughly eight to sixteen gigabytes available. More memory allows larger models.
A dedicated graphics card makes it fast. A plain processor works and is slower – think tens of seconds rather than instant.
How to start
Install one of the local runner applications, which handle downloading and running models through a simple interface. Pull a mid-sized general model. Ask it something.
That is the whole setup.
What local models do well
Summarising, extracting structured data from text, classification, drafting, translation of common languages, and routine code assistance.
Anything where the task is well-defined and the output is checkable.
What they do less well
Broad factual knowledge, where a smaller model knows less and invents more.
Complex reasoning across many steps.
Less common languages, where quality drops noticeably.
Why bother
Nothing leaves your machine, which resolves confidentiality concerns entirely. No usage limits, no per-request cost, no service outage, and no terms changing under you.
The honest comparison
A local model is meaningfully worse than the best commercial service at hard tasks and entirely adequate at most routine ones.
The practical arrangement for most people is both – local for volume and anything sensitive, a commercial service for the cases where quality justifies it.