An Italian AI Search Engine That Runs on One Gaming PC

An Italian researcher has released ItColBERT, an AI search system built specifically for the Italian language. Until now, Italian search has been handled almost entirely by giant multilingual models that juggle dozens of languages at once — a bit like hiring one translator for 50 languages instead of a native speaker. ItColBERT is small: 135 million internal settings (parameters), trained in roughly 14.5 hours on a single RTX 3090, the kind of graphics card a gamer might own.

It works. Across four Italian search tests, it outperformed every general-purpose competitor except one, called mLateOn — while using 2 to 4.4 times fewer settings. But the headline finding is counterintuitive. The single biggest improvement didn't come from more training; it came from a simple trick applied at search time: breaking documents into smaller pieces before feeding them in. That alone lifted the search-quality score by 0.0602 points, a statistically solid gain. Two other ideas — mining harder practice examples and training on longer passages — both failed their pre-set tests.

Why should you care? If you speak Italian, run a business with Italian customers, or dig through Italian legal, medical or archive documents, search results could get noticeably better — and much cheaper to run. A small model on ordinary hardware costs a fraction of a giant one. The author also released everything openly: the model itself, the code, and the full record of experiments, including the ones that flopped.

The catch: it only handles Italian, it's a research paper rather than an app you can sign up for, and it's not the best on every test. The gains are measured on academic test sets, not your daily Google searches. And the biggest lesson is oddly humble — a plain setup trick beat extra training, a reminder that bigger AI isn't always smarter AI.