This Tiny AI Sees Like a Giant — and Could Fit in Your Doorbell

Most AI that "sees" — spotting a pedestrian, a package, a broken part on a factory line — is enormous. These models need powerful chips or a round trip to a data center, which costs money, drains batteries, and means your camera footage may leave your home. A small team at Comexp Research Lab in Nizhniy Novgorod, Russia, is trying a different route. Instead of feeding the AI raw pixels, they first describe the scene in a structured way, then let a much smaller network do the recognizing.

The system is called TAPe+ML v3, and it handles three jobs at once: labeling what's in an image, finding where objects are, and tracing their outlines. The headline number is size: fewer than 100,000 internal settings, or "parameters." Big commercial vision models often use hundreds of millions. On COCO, a widely used public test set, it scored 84.7 for spotting objects and 80.7 for outlining them. The authors say this shows you can shift the heavy lifting from a giant network to a smarter way of describing the picture.

Here's the honest catch. This is a preprint — a self-published paper that hasn't been checked by other scientists — and the results come from one team with no independent repeat yet. Claims this strong from a model this small are unusual, and the comparison totals depend heavily on how the test was run. Also, the clever scene-describing step isn't free; it may simply move the computing cost somewhere else.

The practical upside, if it survives scrutiny: cameras that run AI on the device itself. No cloud subscription, no waiting, no uploading your living room to a server. Factory and medical imaging could get cheaper sensors with smarter software. And because the method is described as needing less training data, the expensive human labeling that props up today's vision AI might shrink.