Karpo-Decision-31B tops open source decision model leaderboard
The readout-head design removes generation entirely, meaning every inference returns a valid option with a fixed latency profile — a structural difference from generative classifiers that teams can exploit for high-volume ranking pipelines. Teams running retrieval, relevance scoring, or agent decision loops can deploy the open weights and tune the act-or-ask threshold on their own data. Against Jev 1.14, the halved median latency and ~20% lower per-decision cost give it a concrete total-cost edge at scale.