Jev introduces a new shape of LLM

TypeSafe AI’s Jev is a new class of model that takes unstructured text or semi-structured “state” objects as input and returns numeric, probabilistic decisions instead of natural language. It supports three query types: binary yes/no (branded Noul, a reference to Bernoulli probabilities) that yields a confidence between 0 and 1; choice queries that return a probability distribution and confidences across provided options; and score queries that place an item on a numeric labeled scale. Jev evaluates many questions in parallel within the context window, is extremely fast, and is priced only on input tokens (their initial model costs about $0.042 per million tokens, cheaper than GPT-5 Nano), with output effectively free.

The framing as a “decision” model makes it well suited for classification tasks - spam filtering, label suggestion, prioritization, ranking and search reranking by scoring candidate results - and its low cost enables large-scale evaluation experiments. It also reintroduces opaque, black-box behavior because users receive only numeric outputs without provenance or explanations, raising bias and auditability concerns for sensitive uses like hiring. Early community experiments have repurposed Jev for playful uses (a chat-by-prediction hack, left-pad, a 2048 player), and open-weight projects are attempting similar decision models and benchmarks (Kev on Qwen 3.5, a JevBench) to reproduce and compare performance.