Decision AI models return probabilities for software-controlled choices
Decision AI models are designed to return typed answers rather than generated paragraphs. Given text and predefined questions, they can produce a choice, a score or a probability that a statement is true. The results are intended for routing, classification, triage, guardrails, reranking and evaluation, where software can act on bounded outputs. They are not intended for summaries, explanations, exact arithmetic or open-ended writing.
TypeSafe’s Jev, launched after two years in stealth, supports Choice, Score and Noul questions, with up to 255 options for Choice. TypeSafe reports response times of 70–500 milliseconds. Its price is $0.042 per million input tokens, while output is free. In TypeSafe’s workflow tests, Jev recorded 67.8% mean accuracy, matching Claude Sonnet 5, but the best comparison model reached 74.1%.
Fastino Labs offers the hosted GLiDE and the Apache 2.0 open-weight GLiNER2.5-Decide. Other local options include Laya, JevK5, OpenJev and kev. Reported results come from different test suites and are not directly comparable. Jev and GLiDE are API-only, while the open-weight systems can run locally. The category is being adopted for agent control, safety checks and low-latency applications.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.