Jev turns LLMs into decision models

TypeSafe AI has unveiled Jev, a model format it calls “System One models,” with “decision models” presented as a clearer description. Jev accepts text or semi-structured inputs, then returns floating point values for categories, yes/no judgments, ratings, and confidence scores rather than generating text.

The service is designed around a “state” object, such as an article, customer profile, or other record, paired with one or more questions. It supports yes/no “Noul” questions, choice questions that return confidence and a probability distribution, and score questions that place an answer along a described numeric range. Questions are evaluated in parallel, making the format suited to spam detection, labeling, prioritization, ranking, and search reranking.

Pricing is positioned as a major differentiator: Jev charges only for input, with output free, and its first model costs $0.042 per million tokens, below OpenAI’s GPT-5 Nano at $0.05/million. The approach also deepens black-box concerns because decisions are returned as numbers without explanations, making bias testing and structured evaluations especially important.

Early experimentation has already produced unusual demos, including chat, left-pad, and 2048 implementations. Open-weight recreations are also emerging, including Kev, which uses Qwen 3.5 to produce 0.8B, 4B, and 9B models, alongside a JevBench benchmark for Jev-class decision models.