TypeSafe AI launches Jev, a model that answers in numbers, not text
TypeSafe AI last week unveiled Jev, the first example of a category it calls System One models. Developer Simon Willison, who wrote up the release, says he prefers the term decision models, a name he credits to Maggie Appleton. Jev still takes text as input, but instead of generating text back it returns floating point numbers: confidence scores tied to categories, yes or no questions, ratings, and probability distributions. TypeSafe describes it as a frontier intelligence function call: unstructured state in, typed probabilistic decisions out.
To use Jev, a developer sends a state object, a string, an array of strings, or a set of name value pairs describing an article, a customer, or any other record, along with one or more questions. Three question types are supported: yes or no questions, which Jev calls Noul questions (TypeSafe's CEO confirmed on Hacker News that the name is short for Bernoulli, from the Bernoulli distribution), returning a confidence score between 0 and 1; choice questions, where the model picks from a set of options and returns a probability distribution across them; and score questions, where the model places a value along a described numeric range. Questions on the same state are evaluated in parallel, so sending many questions takes about as long as sending one.
Pricing charges only for input: $0.042 per million tokens, with output free, which TypeSafe says undercuts OpenAI's GPT 5 Nano at $0.05 per million. Willison sees the model as suited to classification style tasks such as spam detection, label suggestion, prioritization, and ranking. He has used it himself for search reranking: fetch 100 likely candidates with a cheap algorithm like BM25, then have Jev score all 100 for relevance against the original query.
Willison also flags a downside. Jev is, in his view, a step further into black box territory than an ordinary LLM: a regular model can at least be asked to explain its answer, even if the explanation cannot be trusted, while Jev returns only a number with nothing to inspect. He argues this puts bias concerns front and center and says he hopes nobody uses Jev to rank job applicants, since a single floating point score could hide bias that would be hard to pick apart. In his own test, he had Jev answer a yes or no Good city? question for every city in the San Francisco Bay Area; it rated Cupertino highest and East Palo Alto lowest. He concludes that evals and structured experiments matter even more for a model like this, though Jev's low cost makes running large batches of them cheap.
The release has drawn fast community experimentation. Kyle Pena built jevchat, which turns Jev into a chat model by repeatedly asking it which symbol should come next; Fatih Kadir Akın built jev-leftpad, which implements left-pad by asking Jev how many spaces are needed to reach a target length; and Andy Gayton built jev-2048, which uses Jev to play the sliding puzzle game 2048. Separately, an open weight recreation called Kev has appeared, built on top of Qwen 3.5 and offered in 0.8B, 4B, and 9B sizes, alongside a community benchmark called JevBench for comparing Jev class decision models.
Key facts
- TypeSafe AI unveiled Jev, the first System One (or decision) model: it takes text input but returns floating point numbers, not text, for yes/no (Noul), choice, and score questions.
- Jev is priced at $0.042 per million input tokens with free output, which TypeSafe says beats OpenAI's GPT-5 Nano at $0.05 per million; questions on the same input run in parallel.
- Simon Willison used it for search reranking (BM25 fetches 100 candidates, Jev scores them for relevance) and sees it fitting classification tasks like spam detection and prioritization.
- Willison warns Jev is more of a black box than an ordinary LLM and could hide bias; in his own test it rated Cupertino the best and East Palo Alto the worst Bay Area city on a yes/no question.
- Within days, developers built jevchat, jev-leftpad and jev-2048 on top of Jev, and an open-weight clone called Kev (built on Qwen 3.5, in 0.8B/4B/9B sizes) plus a JevBench benchmark appeared.
Why it matters
Jev is not another chatbot: it is a new shape of LLM output. Where a normal model answers in text, Jev answers only in numbers, confidence scores, probability distributions and ratings, tied to whatever question the caller sends alongside a block of text. Because there is no text to generate, TypeSafe charges nothing for output and very little for input, which makes the model cheap and fast for the kind of yes/no or scoring calls that classification pipelines run by the thousand.
Who it affects
It is aimed at developers building classification, ranking, moderation or search systems who currently prompt a regular LLM for a label or a score and throw away everything except one word or number. Willison's own use, reranking search results, is a direct example. It also concerns anyone building or relying on scoring systems that touch people, since he singles out ranking job applicants as a use he hopes nobody makes.
How to use it
A caller sends a state object, a string, a list of strings, or name-value pairs describing a record, together with one or more questions. Noul questions ask yes or no and return a confidence score from 0 to 1; choice questions return a probability distribution across a set of options; score questions return a value along a described numeric range. Multiple questions against the same state are evaluated in parallel. Input costs $0.042 per million tokens; output is free.
How solid is it
The account comes from Simon Willison's own hands-on testing, including a search-reranking experiment and a city-ranking test, published about a week after Jev's release. Independent confirmation comes from the speed of community uptake: several small projects (jevchat, jev-leftpad, jev-2048) and an open-weight recreation, Kev, plus a community benchmark called JevBench, appeared within days, which suggests the API behaves as described, though no formal third-party benchmark of Jev itself is cited.
Risks and caveats
Willison's central worry is opacity: a regular LLM can at least be asked to justify an answer, however unreliably, while Jev returns a bare number with nothing to inspect. He argues this makes bias harder to catch, not easier, and points to his own experiment, where Jev rated Cupertino the best and East Palo Alto the worst Bay Area city on a plain yes/no question, as a sign that unexamined bias can surface without explanation. His conclusion is that evals and structured testing matter more here than for ordinary LLM use, even if Jev's low price makes running them cheap.
“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
— TypeSafe AI, describing Jev