Small AI Models Aren't Bad at Databases — Their Tools Are

Researchers at a Turkish university ran an unusual experiment: for eleven days they let 39 small, free AI models — the kind you can download and run on your own laptop — actually operate a database program. Not just answer questions about databases, but take actions inside it: open tables, run queries, write files. They logged 8,199 attempts across six different kinds of tasks. The common assumption is that models this small simply lack the brainpower for that kind of work. The results say otherwise.

Of roughly 2,100 failures tied to a specific model, about 76% happened in runs where the AI had already started using the tools. The biggest single category, 36%, was what the researchers call "transport": the AI reached for the right tool but the request never made it through to a finished result. The smallest category — the AI never touching a tool at all, which you'd expect if it were simply too weak — was just 17%. In other words, the bottleneck was mostly the road, not the driver.

Here's the strange part. The system's logs recorded that a request had been rejected, but never recorded what the AI actually asked for. So for ten days, nobody could see the cause. Once the team started capturing those details, they found five outright bugs in the server software — including one where a tool demanded a field that a companion tool was forbidden from supplying, then failed the run for the missing field. Five fixes to the server, touching no AI model and no prompts, improved six models by 6 to 21 tasks out of 30.

The paper also flags a sneaky measurement trap. One 7.1 GB model, given no memory limit, loaded itself at its maximum size and ate 51 GB of a 64 GB machine. In any normal log, that looks identical to the AI just being slow. If published AI test scores were collected this way, some of them may be measuring hardware, not intelligence. The team released their data and a checker that reproduces every number.