MotherDuck's prompt_jev(): SQL text classification 50x faster at 1% of LLM cost

dbreunig · x · 2026-09-22

MotherDuck shipped promptjev(), a SQL function integrating Jev, a new model from TypeSafe AI. On a 100,000-row text classification benchmark, Jev matched frontier-LLM accuracy in 40 seconds for $0.50, versus over 30 minutes and $37 for the LLM — roughly 50x faster at 1% of the cost.

Jev is pitched as a "frontier-intelligence function call": unstructured text in, typed probabilistic decisions out (label, score, or yes/no with confidence). As a SQL scalar function, its output needs no parsing and can be filtered, joined, and aggregated in the same statement. Compared to training your own BERT (labeled data + model maintenance) or running an LLM over whole tables (slow, expensive), this speed/cost point makes previously uneconomical table-scale workloads viable. Users are already asking for a plain DuckDB extension.

Related event: MotherDuck Integrates Jev: 50x Faster SQL Text Classification at 1% Cost(4 posts)→

Original post →

More from Apps

Apps channel →