Simon Willison has spent a week with Jev and describes it as a new shape of language model. TypeSafe AI calls the category System One models; Willison, alongside Maggie Appleton, prefers decision models. I wrote about the model’s launch last week.
The difference is in the answer. Jev takes text and gives back no text at all — floating point numbers instead: categories, yes/no probabilities, ratings, each with a confidence score. TypeSafe describes it as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.
One state, as many questions as you like
You build a state object — a string, an array of strings, or a set of name-value pairs. It might be an article, a customer, any kind of record. Then you send one question with it, or a hundred. Questions are evaluated in parallel, so a hundred take about as long as one.
There are three question types. Noul questions are yes/no: you state something and get back a number between 0 and 1. Choice questions have the model pick from options you supply and hand you the full distribution. Score questions give it a described scale and get a point along it.
Pricing is the other oddity. Only input is billed, output is free, and input runs at $0.042 per million tokens. That undercuts GPT-5 Nano at $0.05.
What Willison uses it for
Anything that can be phrased as classification: spam detection, label suggestions, prioritization, ranking. His most interesting example is search reranking. Pull a hundred candidates with something cheap like BM25, then have Jev score each of them against the original query.
The community got stranger faster. Kyle Pena’s jevchat turns Jev into a chat model by asking it one question at every step: given the question and the reply so far, which symbol comes next? A commenter on Hacker News called it the digital equivalent of Morty talking to the death crystal. There is also jev-leftpad by Fatih Kadir Akın and jev-2048 by Andy Gayton.
A week old and already cloned
Open-weight recreations are appearing in parallel. Kev builds on Qwen 3.5 and comes in 0.8B, 4B and 9B sizes. A JevBench for comparing models of this class already exists.
For something released just under a week ago, that is a lot of traffic — and a decent signal that this filled a gap plenty of people had felt.
The black box gets blacker
Willison’s objection is the most valuable part of his post. An LLM is already a black box, but you can at least ask it to justify itself, even if the justification may be nonsense. Jev hands you a number. If it flags something as spam, you never learn which signal tipped it.
Bias is where that gets serious. Willison says outright that he hopes nobody points Jev at job applications. His own test: he scored every city in the Bay Area on the question «Good city?» Cupertino came top, East Palo Alto bottom.
Which is why decision models live and die by evals. The price is the good news there: running a thousand test prompts costs a few cents. A tool whose only window inward is the test had better be cheap enough to test often.