You do not need a gaming rig to run a language model locally. This guide was written and tested on a 2019 laptop with 8GB RAM and no GPU worth mentioning. It is not fast. It works.
Step 1 — Install Ollama
Download the installer from the official site and run it. On Windows it registers a background service; you will interact with it through a terminal.
ollama --versionIf that prints a version number, the service is running.
Step 2 — Pick a model that fits
This is where most 8GB guides lie to you. Forget the 7B models — after the OS takes its share you have maybe 5GB to play with. These are the ones we ran successfully:
| Model | RAM used | Speed on our machine |
|---|---|---|
| qwen2.5:3b | ~3.3GB | Usable, ~8 tok/s |
| llama3.2:3b | ~3.4GB | Usable, ~7 tok/s |
| phi3:mini | ~3.8GB | Slower but sharper answers |
Step 3 — Run it
ollama run qwen2.5:3bFirst run downloads the model (2–3GB, be patient). After that you get a chat prompt in your terminal.
The catch: if you open a browser with 20 tabs while the model is loaded, Windows starts swapping to disk and everything — the model, the browser, the mouse cursor — dies together. Close what you can before long sessions.
Troubleshooting
Model loads then answers gibberish — you likely pulled a quantisation too aggressive for the model; delete and re-pull the default tag. connection refused on the API port — the background service is not running; start Ollama from the Start menu first.