Decision AI Models Emerge: Jev, GLiDE, GLiNER2.5-Decide

Decision AI models, a novel category of artificial intelligence, are designed to return structured decisions rather than free-form text. These models accept text input along with typed questions and output choices, scores, or binary probabilities that can be directly integrated into code for branching logic. The category gained significant attention with the launch of Jev by TypeSafe AI, which had been in stealth development for two years. Within three weeks of Jev's release, Fastino Labs introduced two competing models, and open-source developers published several reproductions of Jev's architecture. This development signifies a rapid evolution in the decision-making AI landscape.

Jev operates by accepting a 'state,' which can be a string, array, or set of name-value pairs, and one or more questions. According to TypeSafe's documentation, Jev supports three primitive output types: 'Choice,' allowing the selection of one option from a list with associated probabilities and confidence levels, supporting up to 255 options; 'Score,' which rates the state on a rubric of ordered levels with probabilities and confidence; and 'Noul' (short for Bernoulli), providing a 0 to 1 probability that a given statement is true. TypeSafe emphasizes that each question is evaluated in parallel and in isolation against the same state, meaning the addition of more questions minimally impacts response time. A key architectural advantage highlighted by TypeSafe is that Jev never generates strings, thus preventing type errors in downstream code. The underlying architecture is described as a novel design featuring a parallel sampler and a training methodology termed Reinforcement Learning for Calibrated Decisions (RLCD). Unlike Reinforcement Learning from Human Feedback (RLHF), which optimizes for human preference, RLCD specifically targets calibrated probabilities, ensuring that higher confidence scores correlate with greater accuracy.

Pricing is a critical factor in the adoption of these decision AI models. TypeSafe's Jev is priced at $0.042 per million input tokens, with output tokens being free. OpenRouter lists Jev as having a 32K context window. TypeSafe reports end-to-end response times for Jev ranging from 70 to 500 milliseconds. The emergence of these models suggests a shift towards AI systems that provide more deterministic and actionable outputs, facilitating integration into automated workflows and complex decision-making processes across various industries. The rapid response from competitors and the open-source community indicates a strong interest in this new paradigm of AI.