Jev: What TypeSafe's Decision Model Is, and Which Page You Actually Need

Jev is a decision model made by TypeSafe AI: you send it a piece of state and a set of named questions, and it returns a typed answer for each one — a label from a list you supplied, a score on a rubric you defined, or a yes/no claim with a probability attached — rather than writing a sentence back. The model in service today is Jev 1.13, whose vendor versioned ID is jev-1.13.0; it belongs to a model class TypeSafe AI calls System One; and the lab behind it is TypeSafe AI, a small San Francisco company that publishes at typesafe.ai. On OrcaRouter the exact string is typesafe/jev-1.13. Those are four separate things — a model, a version, a class, and a company — and a large share of the people arriving on this page are looking for one of them without being sure which. A good number of them arrive from a search typed in Chinese: the queries jev模型 and jev 模型 together carry more than nine thousand of the impressions we measure for this name, and what those readers want is the identifier rather than a translation. There is no Chinese name for any of it. Jev, Jev 1.13, jev-1.13.0 and typesafe/jev-1.13 are written the same way in every language, and the strings below are the ones to copy.

This page is the reference entry under the name Jev. It is not a launch post: Jev 1.13 first appeared publicly on 2026-09-15, which puts it well outside the seven-day window this blog writes news to, and TypeSafe AI has never published a per-post date on its own blog, so a single-day "Jev release date" quoted at you from anywhere should be read as somebody's snapshot rather than the vendor announcing anything. What this page is instead is the index. The other Jev pages already on this site each answer one narrow question; this is the one that tells you what Jev is, settles the naming, and points you at the right page for whatever you actually came to find out.

The names, and why so many people think they are three products

TypeSafe AI is the company. It builds what it describes as machine-native intelligence infrastructure for automation, was founded by Diogo Almeida (CEO) with Sasha Sheng (COO) and Erik Gafni (CTO), and came out of two years in stealth with a Series A led by Andreessen Horowitz alongside Sequoia Capital and DCVC. Its own site carries the line "We're building prod, not God." When a reader types the single word "typesafe" and means the AI model, this is the entity they mean — not the programming-language adjective that the same word spells.

System One is the class. TypeSafe AI introduced it as a category of model built to make fast, structured decisions that software can consume directly, borrowing the name from the System 1 / System 2 split in behavioural economics: the fast, pattern-matching half of cognition. The class definition is deliberately narrow. System One models do not write replies, do not produce code and do not generate explanations of their reasoning, and as of the current documentation the input side is text only. Jev 1.13 is TypeSafe AI's flagship model and the first member of that class; nothing else TypeSafe AI ships is a System One model today.

Jev is the product. It is the thing you call, and the thing this page, the other pages under this name and every comparison in this stem are about. The version is part of the product's identity rather than a footnote, which is where most of the confusion starts, so it is worth naming the strings exactly:

• jev-1.13.0 — the vendor's versioned model ID, the value the API returns in the response's model field to tell you which build answered you.

• jev-latest — a TypeSafe AI alias meaning the most recent stable official release. It resolves to jev-1.13.0 today, and it is the SDK default.

• jev-preview — a second alias meaning the most recent release whether or not it is official. It also resolves to jev-1.13.0 right now, because no preview build exists; the documentation says so plainly.

• typesafe/jev-1.13 — the catalogue ID on our side, and the one to copy if you are calling through OrcaRouter.

• Jev 1.13 — the name to use in prose and in search boxes. "Jev 1.13" and "Jev" are not two models.

There is one more name that catches people, and it is a trap rather than a relation. RSI-Jev is a small 4B vision-language model published on 2026-10-06 under Apache-2.0 by researchers unconnected to TypeSafe AI, and its own card states that it is not affiliated with TypeSafe AI. If you searched for Jev, found a downloadable checkpoint and wondered why TypeSafe AI also sells an API for it, that is why: they are two different things sharing three letters. The full breakdown of the two, including what the fork can and cannot do, is in RSI-Jev vs Jev 1.13: One You Download, One You Call, and the identity question on its own is in What Is RSI-Jev?.

If what you wanted was not the naming but the model's behaviour, the page for that is jev-model-explained, which covers why the answers come back as labels and what that does to a prompt.

What Jev 1.13 is built for

Jev 1.13 is built for decisions that are already closed-form before the model runs. You either supply the list of labels, or you define the scale, or you state the claim; the model's job is to pick, score or attest, and to attach a probability to its answer. Every question in a request is evaluated in parallel against a single shared read of the state rather than one after another, which is why attaching twenty questions to one request costs barely more time than attaching one — and why there are no output tokens to meter at all. The documented fits are the unglamorous ones that run at volume: routing and classification, guardrail checks, per-turn verification inside an agent loop, and scoring large sets of documents against a fixed rubric.

What it does not do is write. It produces no prose, no code and no explanation of its reasoning, and it takes no image or audio input. The shape is the other constraint: the documented budget is 64K tokens per request, with 32K of that available to the state alongside its longest question — generous for a piece of state plus its questions, but not a long-context model, so a workflow that depends on reading a whole contract in one pass is a workflow Jev 1.13 cannot take. Beyond the shape, TypeSafe AI maintains its own list of the ways the model gets decisions wrong — reading questions literally, counting and arithmetic, date and time comparison, indirection, states that are large and mostly irrelevant, adversarial content, contradictory instructions, the order of options in a choice, and generation. That list is nine named failure modes with a suggested remedy under each, and it is documented down to the version it applies to. We are not reproducing it here because we already maintain it properly: Where Jev 1.13 Breaks: TypeSafe's Own Limits List, slug jev-limitations, is the page to read before you put this behind anything that matters, and it is the page most readers of this hub actually want.

Weights, licence, and what you can download

Jev 1.13 is not open source. TypeSafe AI publishes no model weights for it and no licence that would let you run it yourself; what the company does open-source is tooling — client SDKs and helpers under MIT and Apache-2.0 — while the model stays a hosted API. On the data side the published commitments are a promise not to train models on customer data and zero data retention for enterprise customers. So there is no "which quantisation do I download" answer for Jev 1.13, and no GGUF, and anyone telling you otherwise is describing the other thing with Jev in its name.

The downloadable artefact in this corner of the market is RSI-Jev v6.0-VL, the 4B Apache-2.0 model mentioned above, which you can hold on your own disk and serve yourself. It is a genuine option if your reason for wanting open weights is that the weights must be yours, and it is a reimplementation of the same idea rather than a release of Jev 1.13. We argued that out at length in Is Jev Open Source? No, and an Open Third-Party Model Does the Same Job — slug jev-open-source — so treat this section as the pointer and that page as the argument. The one-line version: what TypeSafe AI's licence and published artefacts permit is calling Jev 1.13 through the vendor's endpoint or through a gateway; what they do not permit, because they do not provide, is running it. If you need weights in hand, the licence you are actually evaluating is RSI-Jev's Apache-2.0, not TypeSafe AI's.

The comparison set

Five head-to-heads exist under this stem, each built around one measured result rather than a general impression. If you are weighing Jev 1.13 against something specific, go straight to the one that matches, because each of them contains a number this page would only flatten:

• jev-vs-claude-fable-5-1 — Jev vs Claude Fable 5.1: The Chess Game That Measured the Wrong Thing. Fable 5.1 built a winning position and lost on the clock; the useful half of the story is what a millisecond latency floor buys you against a model that takes seconds per move.

• jev-vs-deepseek-v4-1-flash — Jev vs D​eepSeek V4.1 Flash: Eight Cents a Thousand Is Not a Moat. On 77 examples of the standard intent-classification set, D​eepSeek V4.1 Flash landed one cent per thousand classifications behind Jev 1.13 and four points ahead on accuracy, which is the honest case against buying a decision model on price.

• jev-vs-gpt-6-astra — Jev vs GPT-6 Astra: The Judge and the Defendant in the Same Comparison. The dashboard that scores Jev 1.13 highest uses the averaged judgments of GPT-6 Astra and Claude Fable 5.1 as its reference answers, so the accused is grading the exam.

• jev-vs-kev — Jev vs Kev: A $95 Fine-Tune That Beats Jev on Support Tickets. Kev is an Apache-2.0 open-weight model that reproduces the Choice / Score / Noul request shape for about $95 of fine-tuning time, and it wins on support-ticket routing while collapsing on harder general evaluations.

• jev-vs-laya — Jev vs Laya: 33 Milliseconds on a T4, and a Random Baseline. Laya is a 421M Apache-2.0 decision model from Convai Innovations that is far more portable than Jev 1.13 and, on its own published zero-shot score, does worse than always answering with the majority class.

Two further comparisons sit alongside those without the "Jev vs" framing, and both are worth knowing about before you conclude you have read everything: openai-decisions-api-first-client-vs-jev, on what happens when the biggest vendor in the market ships a decision API of its own, and rsi-jev-vs-laya, on how the two open-weight alternatives compare once Jev 1.13 is out of the picture. If you want the model's class rather than its rivals, jev-system-one — "System One" as a Model Category — is the page for that, and if you want the earliest reporting we did on Jev before any of this existed, jev-typesafe-system-one-what-we-know is the one to read with the caveats attached to it.

What is not settled

Two things, and we would rather name them than round them off. First, no release date has been announced for the next Jev version — TypeSafe AI has published no version number and no date, and its blog does not carry per-post dates at all, so the only dates we can stand behind are the ones documented elsewhere: a first public build on 2026-09-15, the vendor's removal of the early-access waitlist on 2026-09-21, and the model's arrival in our own catalogue on 2026-09-24. Second, independent evaluation of Jev 1.13 is thin. The accuracy figures most in circulation come from TypeSafe AI's own workflow dashboard, whose reference answers are averaged model judgments rather than ground truth and whose own documentation flags possible harness bias; the model's jaggedness page, which is the vendor's own account of where it fails, was last reviewed on 2026-10-02 and is a more honest document than any benchmark table we have seen for it.

Availability on OrcaRouter

We hold exactly one Jev model in our live catalogue, and it is typesafe/jev-1.13 — spelled with the version, because the version is part of the ID and a reader who copies half of it gets nothing. That single entry is available today. We are not claiming anything about a Jev version history, and we are not claiming any other Jev model: if a page anywhere tells you OrcaRouter serves jev-1.13.0, jev-latest or a Jev 2, the correct string to call is still the one above. It sits behind the same key as everything else we route — 200+ models behind one O​penAI-compatible key at provider list price passed through with 0% markup — with automatic failover across providers, so a decision layer that stops answering does not take the generative path down with it, and so a vendor price move lands on our side the same day rather than at the next billing cycle.

Which page you want

If you came here from a search box rather than from a link, the odds are good that this hub is not the page you needed and one of the ten below is. It is a short list on purpose.

• If your question is where will this thing get the answer wrong, read jev-limitations. It is the vendor's nine failure modes, checked against what we can see.

• If your question is can I download and run it, read jev-open-source. The short answer is no for Jev 1.13 and yes for the unrelated RSI-Jev, and that page does the work of separating them.

• If your question is should I use this instead of X, read whichever of the five comparisons above names your X. Each one is built on a single measured result, and the one that matters most to your decision is almost certainly in there.

• If your question is what even is a decision model, read jev-system-one, which covers the class rather than this member of it.

• If your question is simply what is Jev, you have it: Jev 1.13, from TypeSafe AI, a System One decision model that returns typed answers instead of writing prose, callable on our side as typesafe/jev-1.13.