Jev is the first model this dataset had no category for

Every model in this dataset is filed by what it produces. A language model produces text; an image model, an image; a speech model, audio. When TypeSafe AI released Jev on September 15, 2026, none of those fit, because Jev does not produce any of them. It produces decisions.

What Jev does

You give Jev a state — a description of the situation at hand — and a set of typed questions about it, such as a yes-or-no question. It answers every question at once, each with a probability. TypeSafe's announcement puts the difference from a language model plainly: Jev

outputs all probabilities in parallel instead of autoregressively generating by token.

A language model asked the same questions writes its answers out one token at a time, and a program then has to parse the text back into values. Jev skips both steps: the answers arrive as values, with a confidence attached to each. TypeSafe calls this a System One model — borrowing psychology's name for fast, intuitive thinking, popularised by Daniel Kahneman — and its documentation is built around the small judgements software makes all day: routing, scoring, classifying, rather than conversation.

It is served through TypeSafe's API, as jev-1.13.

What it gives up, in its own documentation

TypeSafe publishes a page on where jev-1.13 goes wrong. It says the model struggles with tasks that require numeric precision, can be quite literal in its understanding, and has trouble with extra levels of indirection. Its advice includes keeping arithmetic in code and spelling out each condition exactly.

That is what the trade looks like. A model that answers in one parallel pass has no room to write out intermediate steps, so it suits single judgements and struggles with anything that needs a chain of them.

Why it is not filed as a language model

This dataset's model types answer one question: what does the model produce? Filing Jev as a language model would count a model that emits no text among the LLMs, in every chart that splits the dataset by type. So it has a type of its own, decision, added eight days after it launched.

A new type is only worth adding if it stays narrow, so it has an inclusion test with four parts, all of which must be evidenced by the lab's own documentation (TAXONOMY §6a):

The model's primary output is an answer of a declared type — a choice, a

boolean, a number, a label — not free text.

Each answer comes with a probability, as part of the output.
The answer is not produced by generating text token by token.
The questions and their answer types are supplied at request time.

Each part exists to keep something out. A language model with a JSON mode already returns structured answers, but generates them as tokens — that fails the third. A reranker already scores without generating text, but only answers one fixed question, relevance — that fails the fourth, and so does any classifier trained for a single task.

An imitation within a week

Two days after the launch, an independent developer published Kev: in its own description, a tiny Jev-like family of decision models built on top of Qwen 3.5, which you can train and run yourself. It is not recorded here: it is one person's fine-tune of another lab's model, and a release in this dataset needs a lab behind it. But it shows how quickly the idea was picked up.

What this does not claim

Not that Jev is better than a language model. It does something

different. Its own documentation sends arithmetic back to ordinary code, and a model that produces no text is no help where the output has to be text.

Not that it would stay alone. When this was written, Jev was the only

model in this dataset that passed the test. It did not stay that way for long: within four weeks of its launch, Amazon, Cloudflare, Perplexity, Liquid AI and Microsoft had each released a model that passes it too, starting with Strands Decider 2B on October 1. Most of them were built the way Kev was, on an existing language model, but each has a lab behind it. Whether the approach lasts is still open; it is no longer one company's product.

Not TypeSafe's performance figures. Its speed and cost comparisons appear

in coverage of the launch, but this dataset has not verified them, so they are not repeated here. Jev is also recorded as a milestone, for the event rather than the model.