AI Math Now Runs Perfectly on a Raspberry Pi — No Internet Needed

Small AI models are useful because they can run on your own device — no internet, no data leaving your phone or laptop, no waiting on a server. The catch is that small models are surprisingly bad at things computers should nail: arithmetic, algebra, and step-by-step logic puzzles. Ask them to add two large numbers and they may confidently get it wrong.

This team's fix is almost commonsense. Most "reasoning" questions aren't really reasoning at all — they're structured problems with one exact right answer, so you don't need an AI guessing at them. The researchers built a router (think of it as a smart receptionist) that reads each question and forwards it to the cheapest tool that can solve it correctly. Calculator-style questions go to a deterministic solver, which always gets them right. Only genuinely open-ended word problems reach the small language model.

What makes this notable is how the router was built. Instead of programmers hand-writing the rules, the system learned them automatically, using the small AI model as a helper to test whether a question fits a pattern. The result: a router that got 100% of its sorting decisions right across 100 unseen test questions from three well-known math and logic datasets.

The hardware is the headline. All of this ran on a Raspberry Pi 4B — a credit-card-sized computer with 8 gigabytes of memory and no graphics chip — with no network connection. It scored 98.3% overall, versus 72% for the strongest previous approach, and answered formatted questions in 1 to 11 milliseconds. In its leanest setting it ran 8.8 times faster and used 2.8 times less energy. The catch: it was tested on 100 curated problems, not the messy, ambiguous questions real users ask. And the technique works best for structured tasks — the word problems still leaned on the small model, scoring 93.3%. The paper has been submitted to IEEE for review, so it hasn't yet been independently verified.