Model explainer · independent
What is the Jev model?
Jev is TypeSafe AI's System One decision model. It does not continue text. You give it a piece of state and a set of typed questions, and it returns a constrained value for each question — one option from a list, a number on a scale, or a yes/no — with a probability attached.
That is the whole idea, and it is why searches for "jev model" rarely mean a chat model. The rest of this page covers the names you will see attached to it, what comes back from a call, and the things Jev cannot do.
The short answer
A generative language model predicts the next token until it stops. Jev scores a fixed answer space in one pass. Because the legal answers are declared before the call, there is nothing to parse, no JSON repair step, and no invented enum value for downstream code to trip over.
The practical consequences are a much tighter latency window, an answer you can log as a value, and a probability you can threshold in code. The trade-off is absolute: if the answer is a paragraph, an explanation, or anything you cannot enumerate up front, this is the wrong tool.
Names you will see attached to it
Several labels refer to the same family, and some refer to specific hosted builds. The table separates product names from deployment labels.
| Label | Where it appears | What it means |
|---|---|---|
| System One | TypeSafe documentation | The model category and the API family name |
| Jev | Community usage, press, this directory | The first System One model; the name most people search for |
| Jev 1.13 | Gateway model IDs, third-party tests | A dated hosted build, listed on OpenRouter from 18 September 2026 |
| Jev 1.14 | Gateway model IDs | A later hosted build, listed on OpenRouter from 14 October 2026 |
| ~typesafe/jev-latest | OpenRouter | An alias that redirects to the newest Jev build |
| typesafe/jev-router | OpenRouter | A separate routing entry added 25 September 2026, not a new checkpoint |
Context-window figures differ across sources and dates: earlier materials quoted on this site describe 32k of state within a 64k request budget, while the OpenRouter listing for Jev 1.14 shows 262,144 tokens. These are different builds and different surfaces. Check the current official docs before sizing a request.
What comes back from a call
The request carries a state — text, a document, or structured values — and a map of questions. Each question declares its legal answers. The response carries one typed answer per question.
- →Choice picks one option from a closed set you supply; use it for routing, classification, and tool selection.
- →Score returns a level on a scale you define; use it for urgency, relevance, or quality thresholds.
- →Noul is a two-option gate; use it for approval flags, escalation, and circuit breakers.
- →Multiple questions can be asked against the same state in one request.
const decision = await jev.decide({
state: 'Invoice lists two seats, but only one of us can sign in.',
questions: {
intent: ['refund', 'bug', 'billing', 'other'],
needs_human: 'noul',
urgency: ['score', 1, 5],
},
})
// → typed values with probabilities, no parsing stepWhat the model does not return
No rationale, no prose, no token stream, and no streaming response. If a person needs to read the output, a generative model belongs in that part of the pipeline. The pattern that shows up repeatedly in this directory is both models side by side: generation for anything a human reads, typed decisions for the high-frequency branches that code consumes.
- →Summaries, drafts, chat replies, and code generation are out of scope.
- →Open-ended answers that cannot be enumerated are out of scope.
- →Architecture papers, parameter counts, and weight releases are not published in the vendor materials reviewed here.
Frequently asked questions
Is Jev a large language model?
+
Not in the sense most people mean. It is built for constrained decisions rather than text generation. A generative LLM produces tokens that downstream code must parse; Jev evaluates a declared answer space and returns a typed value. Some materials describe it as a non-autoregressive evaluation over candidate answers, but the architecture and parameter count are not published.
What is Jev 1.14?
+
A dated hosted build of the model. OpenRouter lists typesafe/jev-1.14 as released 14 October 2026 at $0.042 per million input tokens with free output and a 262,144-token context. Treat dated build IDs as deployment labels, not as separate downloadable checkpoints.
Is the Jev model open source?
+
No. TypeSafe has not published weights, an architecture paper, or a training recipe for Jev. Open-weight alternatives in the same category include Cloudflare Clef, Laya, and AWS Strands Decider 2B, all under Apache 2.0.
Why do some searches say "jeb"?
+
In the query data reviewed here it behaves like a spelling variant of Jev. No separate model, project, or product called Jeb is documented in the sources this directory tracks. If you were looking for the decision model, the Jev pages are the right starting point.
Where can I try the Jev model without an API key?
+
This site runs a browser playground. Paste text or upload a document, ask a typed Yes/No, Choice, or Score question, and read the structured answer with the measured round-trip latency for that request. It is a demonstration, not a benchmark.
Keep reading
What is Jev AI?
The conceptual explainer: primitives, good-fit cases, and published pricing.
Jev API reference
Endpoint, authentication, official SDKs, and list pricing.
Try Jev in the browser
Ask a typed question and see the structured answer.
Open source alternatives
Downloadable decision models in the same request shape.
Sources
- TypeSafe AI — official site ↗
- TypeSafe AI — documentation ↗
- OpenRouter — TypeSafe model listings ↗
- OpenRouter — Jev 1.14 model page ↗
Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.