Jev: the LLM that returns a number instead of text

Last week TypeSafe AI unveiled Jev, the first example of a category its makers call "System One models." Simon Willison sides with Maggie Appleton in preferring the name "decision models." Jev is a twist on the usual LLM format: it still takes text as input, but instead of text it returns floating point numbers — categories, yes/no answers, ratings, and their confidence scores.

The company describes Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

Fast and cheap

The model is fast and inexpensive. Regular LLMs bill for input and output tokens, with output usually far pricier. Jev charges only for input — output is free — and the input price of its first model is \(0.042 per million tokens. That's cheaper even than OpenAI's GPT-5 Nano (\)0.05/million).

Three kinds of questions

You compose a "state" object — a string, an array of strings, or a set of name-value pairs describing, say, an article or a customer — and send it to the API alongside your questions. There are three kinds. "Noul" questions (the CEO confirmed on Hacker News that this is short for Bernoulli) pose a statement and return a number between 0 and 1 for how confident the model is that it's true. Choice questions have the model pick one option, with a probability distribution across all of them. Score questions rate along a range of numeric levels with descriptions. Questions run in parallel, so sending dozens takes about as long as sending one.

Jev shines wherever a task can be framed as classification: spam detection, label suggestion, prioritization and ranking. Willison tried it for search reranking — pull 100 candidates with a cheap algorithm like BM25, then have Jev score their relevance.

Black boxes are back in fashion

The uncomfortable part is that Jev pushes even further toward opaque systems. LLMs are already black boxes, but you can at least ask them to justify a decision. Jev doesn't offer even that: feed in all the text you like, and all you get back is a floating point number. If it flags something as spam, you can't tell which signals tipped it off. That makes bias a front-and-center concern — Willison hopes nobody uses it to rank job applicants. In one experiment he had Jev rate San Francisco Bay Area cities on whether each was a "good city"; it put Cupertino top and East Palo Alto bottom.

The takeaway: evals and structured experiments matter even more here than for regular LLMs — and since Jev is so cheap, running thousands of prompts costs just a few cents.

Community fun

Within a week the community produced a flurry of uses: jevchat turns Jev into a (terrible) chat bot by asking at each step which symbol comes next; jev-leftpad reimplements the classic function by asking how many spaces are needed; jev-2048 plays the 2048 puzzle. There have also been open-weight recreations — Kev, built on Qwen 3.5, plus a JevBench benchmark. For a model released less than a week ago, the level of activity is striking.

Source: Simon Willison