Running a model locally has two genuine advantages: nothing leaves your machine, and there is no per-request cost. The constraint is memory, and it is a harder limit than most guides admit.
Memory is the whole story
A model needs to fit in RAM, or in video memory if you have a discrete GPU. As a rough guide, a quantised model needs a little over half a gigabyte per billion parameters. An 8-billion-parameter model wants around 5GB free; a 14-billion one wants closer to 9GB. On a 16GB laptop that is already sharing memory with the browser, the practical ceiling is lower than the specification suggests.
What works well on 16GB
Summarising documents, drafting and rewriting text, classification and tagging, and answering questions about files you provide. These are the tasks where a smaller model is genuinely competitive with a hosted one, because the hard part is reading your material rather than knowing the world.
What does not
Long multi-step reasoning, current-events questions, and anything requiring broad factual recall. Smaller local models are noticeably weaker here, and no amount of prompt work closes the gap. Code generation is a middle case: adequate for small functions, frustrating for anything architectural.
Getting started without a weekend
Install one of the desktop runners that bundle the model download and a chat window, pick a small instruct-tuned model, and try it on your own documents before deciding whether to go further. Watch memory pressure while it runs — if the machine starts swapping, drop to a smaller model rather than waiting it out.
The honest verdict
Local models are a real option for privacy-sensitive document work on ordinary hardware. They are not yet a replacement for a hosted model as a general assistant, and anyone telling you otherwise is testing on a workstation, not a laptop.