Last week TypeSafe AI introduced Jev, which the company presents as the first example of a new model category they call "System One models." Others in the community prefer the term "decision models." Jev departs from the usual large language model (LLM) pattern by still accepting textual inputs but returning floating-point numbers rather than natural language responses. Those numbers represent category choices, yes/no confidences, scalar scores and associated probability distributions.
How Jev works and the supported question types
TypeSafe describes Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out. Users send a "state" object to the API — a string, an array of strings, or name-value pairs describing an article, a customer record, or any other item — and attach one or more questions. Jev returns a numeric response for each question.
There are three question types:
- Noul (yes/no) questions: Jev returns a floating-point value between 0 and 1 that expresses confidence the statement is true. The company’s CEO confirmed on Hacker News that "Noul" is shorthand for Bernoulli, referencing the Bernoulli distribution.
- Choice questions: the model provides a confidence score and a probability distribution across a provided set of options, effectively selecting among them.
- Score questions: given a sequence of numeric levels and descriptions, Jev returns a continuous numeric score somewhere within that range.
The API can accept a single document (state) plus as many questions as fit in the context window; questions are evaluated in parallel so submitting many questions should take roughly the same time as submitting one.
Cost and speed
TypeSafe emphasizes Jev’s speed and low cost. Unlike typical LLM pricing—where both input and output tokens are charged and outputs often cost more—Jev charges only for input tokens and the output is free. Their first model’s input price is $0.042 per million tokens, which the source notes is cheaper than OpenAI’s GPT-5 Nano price of $0.05 per million tokens (used here as a reference).
Appropriate use cases
The decision-model framing helps clarify when Jev is useful: tasks expressible as classification problems. Examples include spam detection, label suggestion, prioritization and ranking. One practical workflow offered is search re-ranking: retrieve ~100 candidate results cheaply using an algorithm like BM25, then have Jev score those candidates for relevance against the original query.
Opacity and bias concerns
Jev increases the degree of black-box behavior relative to many LLM applications. Traditional LLMs are already opaque—you can ask for explanations, but those explanations are not guaranteed to be accurate or useful. Jev gives no textual justification at all, only numeric outputs, so it is harder to trace which content signals influenced a decision. That opacity intensifies concerns about bias: using Jev to rank job applicants, for example, could conceal systemic biases in a single floating-point output and make experimental diagnosis of those biases difficult.
The author tried a small experiment asking Jev to rate cities in the San Francisco Bay Area on whether they are a "Good city?" and reports that Jev ranked Cupertino highest and East Palo Alto lowest—an anecdote that underscores the need for careful evaluation.
Because Jev is inexpensive to call, running many structured evaluations and experiments is affordable: hundreds or thousands of prompts may cost only a few cents, which supports the deeper testing needed to detect undesirable model behavior.
Community experiments and creative uses
Since Jev’s release a number of community-built experiments and playful uses have appeared:
- jevchat (Kyle Pena): converts Jev into a crude chat model by asking at each step "Given the user's question and the reply written so far, which symbol comes next?".
- A Hacker News comment (ericpruitt) compared Jev’s behavior to a pop-culture metaphor: "the digital equivalent of Morty speaking with the death crystal."
- jev-left-pad (Fatih Kadir Akın): implements left-pad by asking Jev how many spaces are needed before a value to reach a target length, using a choice query from 0 to 10 spaces.
- jev-2048 (Andy Gayton): uses Jev to select moves in the 2048 sliding-puzzle game.
Open-weight recreations and benchmarks
There have also been attempts to reproduce Jev-like behavior on open-weight models. Kev is one example that uses Qwen 3.5 to produce 0.8B, 4B, and 9B variants. A community-shared JevBench benchmark has already appeared to compare "Jev-class decision models."
Given that Jev was released less than a week ago, the volume of activity—experiments, creative hacks, and open-weight efforts—is notable.
Conclusion
Jev represents a pragmatic shift toward models optimized for returning typed, probabilistic decisions rather than text. Its speed and low per-input cost make it attractive for many classification and ranking tasks, but the numeric-only outputs aggravate issues of interpretability and bias. The model’s early usage and the rapid emergence of experiments and benchmarks mean that careful, structured evaluation will be essential before Jev-like systems are deployed in high-stakes settings.
