TypeSafe AI's Jev: Decision Models That Return Probabilities, Not Text

Jev is an AI model in a new category the company calls 'System One models' or 'decision models.' It accepts text or semi-structured data and returns floating-point numbers representing probabilities or confidence scores, such as the likelihood a statement is true or a score on a numeric scale.

LLMs are priced on both input and output tokens and return text, while Jev charges only for input ($0.042 per million tokens) with free output and returns probabilistic numbers instead of text. This makes it faster and cheaper for classification tasks, but it offers no textual explanation of its decisions.

It could help organizations processing large text volumes at low cost, such as classifying support tickets, triaging complaints, or reranking internal search results. But caution is needed: Jev is a black box that does not explain its decisions, so rigorous bias testing is essential before any use in decisions affecting individuals, such as hiring or credit.

Community projects like Kev, built on Qwen 3.5 in 0.8B, 4B, and 9B sizes, have emerged, along with a JevBench benchmark for comparing 'decision models.' These alternatives allow local deployment, an attractive option for regional government entities with data-residency restrictions.