Jev introduces a new shape of LLM: System One, aka Decision Models

TypeSafe AI last week unveiled Jev, the first example of a category the company calls "System One models" — a shape Simon Willison says is better named "decision models", the term designer Maggie Appleton also uses.

Jev takes text input like an ordinary large language model but returns no text at all: its output is floating point numbers for categories, yes/no questions and ratings, plus confidence scores. TypeSafe calls it a "frontier-intelligence function call: unstructured state in, typed probabilistic decisions out".

How Jev works

Using the API means composing a "state" object — a string, an array of strings or name-value pairs describing an article, a customer or another record — and sending it with one or more questions. Each question gets its own reply.

Three question types exist. Yes/no questions, which Jev calls "Noul" (the CEO confirmed on Hacker News that the name comes from the Bernoulli distribution), return a number between 0 and 1 for the model's confidence that a statement is true. Choice questions return a confidence score and a probability distribution over the given options; score questions place the document on a described numeric range.

Questions run in parallel, so many of them on one document take roughly as long as one. Pricing is unusual too: only input is charged, at $0.042 per million tokens, and output is free — below OpenAI's GPT-5 Nano at $0.05 per million.

Black boxes and bias

Willison finds the shape of the product uncomfortable: it pushes machine learning further towards opaque black boxes. An ordinary LLM can at least be asked to justify a decision; Jev returns only a number, so nothing explains which content signals put a document in a category. Bias therefore matters more — he hopes nobody uses it to rank job applicants. In one experiment, scoring every San Francisco Bay Area city on whether it is a "Good city?" put Cupertino top and East Palo Alto bottom.

Because Jev is so cheap, he argues evals and structured experiments matter even more than in regular LLM projects: hundreds or thousands of test prompts cost only a few cents.

Community experiments

Unusual projects appeared within days: jevchat by Kyle Pena turns Jev into a poor chatbot asking which symbol comes next, jev-leftpad by Fatih Kadir Akın implements left-pad via a space-count choice question, and jev-2048 by Andy Gayton plays the 2048 puzzle.

Open weight recreations came too, including Kev, which builds 0.8B, 4B and 9B models on Qwen 3.5, and a JevBench benchmark already compares "Jev-class decision models" — striking activity for a release less than a week old.