What is Jev?

Why it is not a language model

An LLM produces tokens; the shape of what comes back is a hope, not a guarantee, which is why applications that use one for decisions end up wrapped in parsing and retries. Jev produces a value of a type you specified. There is nothing to parse and no malformed case to handle — what replaces the retry loop is a branch for when the model is unsure, which is a different and more useful thing to know.

Three shapes, and nothing else

The numbers

TypeSafe publishes these; they are the vendor's own figures, not measurements of ours.

Questions

What is a System One model?

A model built for decisions rather than prose. Jev takes one state plus a map of typed questions and returns one typed answer per question — no tokens streamed, nothing to parse, and no way for it to emit a value outside the set you defined. TypeSafe's framing is System One for fast structured judgement, with an ordinary LLM as System Two when you actually need language.

What does confidence mean?

Choice and Score answers carry a confidence value derived from their probability distribution: concentrated means certain, spread out means the model cannot separate the options. It is a second axis — the answer tells you what, confidence tells you whether to act. Noul answers have no confidence field, because the probability itself already carries it: 0.51 is the model saying it does not know.

Can Jev hallucinate?

Not in the shape of its output. A Choice can only ever return one of the options you defined; a Score can only land inside your rubric; a Noul is a number between 0 and 1. It can still be wrong about your text — which is exactly what confidence is for.

Why is it called Jev?

After William Stanley Jevons, the economist. TypeSafe’s launch post says they expect machine intelligence to follow the path coal took after steam engines got efficient — the Jevons paradox, where making something cheaper to use increases how much of it gets used rather than less.

Can I run Jev locally? Is it open source?

No. Jev is a proprietary hosted model reached over TypeSafe’s HTTP API; there are no published weights to download, and the same weights serve every account — it is not fine-tuned or LoRA-adapted per customer. You shape its answers through the request instead: what you put in the state, and how you write each question’s instructions and criteria.

How fast is Jev?

TypeSafe quotes 70–500 ms end to end, against 3–329 seconds for frontier models, and 40–200× faster on the kind of query it is built for. Those are the vendor’s figures. What this site can say from its own logs is that a nine-question request typically comes back in a few hundred milliseconds — every result page shows the measured round trip for that particular call.

What does Jev cost?

$0.042 per million input tokens on jev-1.13.0, with output free — there is no autoregressive decoding to meter. Because the state is read once and every question is evaluated against it in parallel, asking ten things costs roughly what asking one costs. Each result page on this site prints the token count and the cost of that exact call.

How was Jev trained?

With what TypeSafe calls RLCD — Reinforcement Learning for Calibrated Decisions. The stated goal is different from RLHF or RLVR: not answers people prefer or answers that pass a checker, but answers whose probabilities are epistemically honest. That is what makes confidence worth thresholding on rather than decorative.

What are the limits?

Text in only — no images, audio or video. 64k tokens per request, of which the state plus the single longest question must fit in 32k. A Choice takes up to 255 options; a Score takes 2 to 10 levels. Rate limits on early access are 250,000 tokens per second and 1,200 requests per minute, and TypeSafe warns they move.