Jev unveils a novel LLM paradigm — System One, also known as Decision Models

If you build software that needs to make quick judgments about text—spam filtering, labeling, ranking search results—Jev from TypeSafe AI is worth a close look. Released last week, it's the company's first "System One model" (a name I agree with Maggie Appleton should really be "decision model"), and it works very differently from the chatbots most of us are used to. This article explains what Jev actually does, how much it costs, where it shines, and why its black-box nature deserves some healthy skepticism.

What Exactly Is Jev?

Jev looks like a large language model on the input side: you feed it text. But on the output side, everything changes. Instead of words, it returns floating point numbers—confidence scores tied to categories, yes/no answers, ratings, and probability distributions.

TypeSafe's own description captures the idea neatly: think of Jev as "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out."

Speed and price are the headline features here. Standard LLMs bill you for both input and output tokens, and output tokens usually cost noticeably more. Jev flips that around entirely: output is free, and you only pay for input. The first model runs at $0.042 per million tokens—undercutting even OpenAI's GPT-5 Nano, which sits at $0.05 per million.

How the API Works

Using Jev starts with building a "state" object. This can be a plain string, an array of strings, or a set of name-value pairs—essentially a description of whatever you're asking about: an article, a customer record, anything text-shaped.

You send that state to the API along with one or more questions. There are three question types:

1. Yes/No questions — Jev calls these "Noul" questions. The CEO confirmed on Hacker News that this is short for Bernoulli, as in the Bernoulli distribution. You state a claim and receive a number between 0 and 1 representing how confident the model is that it's true.

1. Choice questions — the model picks from options you supply, returning both a confidence score and a probability distribution spread across every option.

1. Score questions — you provide a sequence of numeric levels with descriptions, and Jev returns a floating point score somewhere along that range.

One nice engineering detail: you can attach many questions to a single document, limited only by the context window. Questions run in parallel, so batching twenty of them takes roughly as long as sending one.

Where Jev Makes Sense

The "decision model" framing is the best guide to picking use cases. Anything you can phrase as classification is a natural fit—spam detection, suggesting labels, prioritizing tasks, ranking items.

Search reranking is another promising pattern. Fetch 100 likely matches using a cheap algorithm such as BM25, then let Jev score all 100 candidates for relevance against the original query. That's a task that would be overkill for a full LLM but lands squarely in Jev's sweet spot.

Black Boxes Are Back in Fashion

Here's the part that makes me uneasy: Jev represents a step backward in transparency, and it's worth being honest about that.

LLMs are already black boxes. You can at least ask one to explain its reasoning, even if there's no guarantee that explanation is accurate or useful. Jev doesn't offer even that much. Feed it all the text you like, and the only thing coming back is a number. If Jev flags an email as spam, which signals triggered the verdict? You have no way to know.

That opacity puts bias concerns front and center. I sincerely hope nobody uses Jev to rank job applicants—a single floating point number could hide all kinds of baked-in bias, and teasing that bias out experimentally would be genuinely difficult.

I ran one small experiment myself: I asked Jev to answer yes/no on whether each city in the San Francisco Bay Area was a "Good city?" It ranked Cupertino at the top and East Palo Alto at the bottom. Huh.

The practical takeaway is that evals and structured experiments matter even more here than with ordinary LLM projects. The silver lining: Jev is cheap enough that running hundreds or even thousands of experimental prompts costs only a few cents.

Unconventional Uses the Community Found

Watching people stretch Jev in weird directions over the past few days has been genuinely entertaining. A few favorites:

• jevchat by Kyle Pena turns Jev into a (terrible) chat model. At each step it asks one question: given the user's question and the reply written so far, which symbol comes next? As ericpruitt put it on Hacker News, "It's the digital equivalent of Morty speaking with the death crystal."
• jev-leftpad by Fatih Kadir Akın implements left-pad with the prompt "How many spaces are needed before value to reach targetLength?" plus a choice query offering options from "0 spaces are needed" to "10 spaces are needed."
• jev-2048 by Andy Gayton uses Jev to play the 2048 sliding puzzle game.

Open Weight Recreations Are Already Here

A wave of projects is also trying to recreate Jev on top of open weight models. One notable example is Kev, built on Qwen 3.5, which ships in 0.8B, 4B, and 9B sizes. The accompanying Hacker News thread even surfaced a JevBench benchmark that's already appeared to compare "Jev-class decision models" against each other.

Considering Jev launched less than a week ago, the amount of surrounding activity is seriously impressive.

Who should use it / who should skip it: If your product needs fast, cheap classification, scoring, or reranking at scale, and you're willing to invest in proper evals, Jev looks like a strong fit. If you need explainable decisions, auditable reasoning, or you're working in high-stakes domains where bias must be surfaced and documented, this is probably not your tool—yet.