I put a local AI model on my five-year-old laptop. It's slower than the cloud and I keep using it
No internet, no subscription, no data leaving the house. The responses take a breath longer and I've stopped caring — here's where local actually fits.
The experiment was supposed to be a dismissal. Five-year-old laptop, 16GB of RAM, no discrete GPU to speak of — run a local model and confirm what everyone says: nice toy, unusable for real work. A week later the model is still installed and I keep opening it. Not for everything. For more than I expected.
The numbers first, because you'll ask: a mid-size open model, quantized, generates at roughly speaking-pace on this machine. First token takes a few seconds. That's the whole horror story, and it's real — for anything long-form, the cloud is five to ten times faster and smarter. I'm not going to pretend the local model wrote my newsletter.
What it does instead is three jobs where 'slower but private' wins outright. Meeting notes with anything client-identifying: the transcript never leaves the machine, which my lawyer friends asked about and my cloud tools had no good answer for. Second-draft reading of my own writing at midnight when ideas are thin and the internet is off. And the simple joy of asking dumb questions with zero logging anxiety.
The surprise was reliability, not capability. The cloud tools have outages, rate limits, and this week's price restructuring. The local model has the same personality every single day. It's a bad baseline for cutting-edge work and a great baseline for the boring middle of your workload — the stuff you'd never brag about but do constantly.
Who should skip this: anyone whose questions need frontier-model intelligence, and anyone without patience for a download and a settings page. Who should try it: anyone whose list of 'things I can't paste into a chatbot' is getting longer than the things they can. The laptop didn't become a supercomputer. It became a place where the sensitive half of my questions has somewhere to go.