What is Jev?
Jev is TypeSafe AI's flagship model and the first "System One" model: you send it some text, called the state, plus typed questions, and it returns structured answers with probabilities instead of generated text (TypeSafe docs). It is trained with RLCD, Reinforcement Learning for Calibrated Decisions, so your code can use those probabilities to decide when to act and when to ask a person.
Jev in plain terms
An LLM writes text for people to read. Jev does not write anything. You give it a state, such as a support ticket, a record, or a chat log, and a set of questions whose possible answers you define in advance. It returns a typed value for each question, plus a probability for each possible answer (System One).
All questions in one request are answered in parallel and independently against the same state, so one answer does not leak into another (Introduction). TypeSafe says most queries complete in about 100 ms (How to build with TypeSafe).
Jev accepts text only: a string, a JSON object, or an array of text. Images, audio, and video are not supported yet. English is its primary training language; other languages work but less well (Models).
The three question types
TypeSafe calls these question types "primitives." You can mix all three in one call (Primitives).
| Type | What it asks | What comes back |
|---|---|---|
| Choice | Pick one option from a list you define (up to 255 options) | choice, probabilities, confidence |
| Score | Rate the state on an ordered rubric (2 to 10 levels) | score, legend, probabilities, confidence |
| Noul | Is this yes/no statement true? | noul, a number from 0 (no) to 1 (yes) |
The option and level limits come from the API reference. A support queue might ask which team owns a ticket (Choice), how frustrated the customer is (Score), and whether it is urgent (Noul), all in one request.
Where RLCD and calibration fit
TypeSafe describes three ways to post-train a language model. RLHF trains models to produce responses people prefer, which gave us chatbots. RLVR trains reasoning models on verifiable rewards. RLCD trains Jev to return decisions with calibrated probabilities instead of text (AI primer).
"Calibrated" means that across many predictions, answers given 0.8 should be right about 80% of the time. The docs are clear that this describes groups of predictions, not a guarantee about any single answer. Choice and Score answers also carry a confidence number from 0 to 1, derived from how spread out the probabilities are (Confidence).
Jev is not fine-tuned on customer data. The same weights serve every account, and you adapt it to your domain through the state and the wording of your questions (Models).
What Jev is good at, and where it struggles
TypeSafe positions Jev for "smart if-statements": classify, route, score, extract, or branch where hand-written rules are too brittle, plus guardrails and verification of LLM outputs (launch post). The docs recommend asking narrow, atomic questions and combining the answers in your own code.
TypeSafe also publishes a list of known weak spots for jev-1.13 (Jev 1.13 jaggedness):
- It reads instructions literally, so vague wording gets a literal answer.
- It does not count or do arithmetic reliably. Keep math in code.
- It reads dates as text, so compare dates in code.
- Double negatives and multi-hop questions cost accuracy.
- Large states full of unrelated detail lower accuracy.
- Injected or adversarial text in the state can move the answer.
- It is not trained to generate text.
Jev is also not a model you can plug into a coding agent. It does not chat, write code, or call tools (Jev with coding agents). For a direct comparison, see Jev vs LLMs.
Versions, status, and origin
TypeSafe announced Jev on September 15, 2026, in early access, after two years in stealth. The post is written by founder Diogo Almeida, who worked on the research behind ChatGPT at OpenAI. Jev is named after the economist William Stanley Jevons, and "System One" comes from Daniel Kahneman's Thinking, Fast and Slow (launch post).
The current model is jev-1.13.0. Both aliases, jev-latest and jev-preview, point to it. An alias can move when a new release ships, so pin the versioned ID if you have tuned thresholds against it (Models).
Where to go next
To see what people have built, browse the use-case directory, starting with Start here and Decision patterns. For cost, read Jev pricing. To make your first call, follow How to use Jev.
Questions
Is Jev an LLM?
Not in the usual sense. Like an LLM it understands natural-language input, but it returns typed decisions and probabilities rather than generated text, and it cannot write replies or code.
Who makes Jev?
TypeSafe AI. Its launch post, published September 15, 2026, is written by founder Diogo Almeida.
Can Jev hallucinate?
TypeSafe says Jev's answers always fit the options you defined, so it cannot invent an answer outside them. It can still pick the wrong option, which is why the docs recommend testing on your own data and routing low-confidence answers to review.
Does Jev accept images or audio?
No. Jev 1.13 accepts text only: strings, JSON objects, or arrays of text. Convert other media to text or structured fields first.
Can I fine-tune Jev on my data?
No. TypeSafe does not fine-tune or LoRA-adapt Jev per customer. You shape its answers through the state and through your question instructions and criteria.
Is Jev trained on my requests?
TypeSafe's docs say Jev is not trained on customer requests or responses. Zero data retention is offered to enterprise customers.
More guides: Jev pricing and API costJev vs LLMs: when to use whichHow to use Jev: a quickstartJev and MCP: using Jev as a Model Context Protocol toolHow to use Jev with Claude CodeJev in an agentic harness: where it fits in an agent loopJev for orchestration: routing requests, models and skillsJev as a judge: evaluating LLM and agent outputsJev for SEO and GEOJev for ads and ad reviewJev for marketing and lead scoringIs there an open source Jev?Can you fine-tune Jev?Designing questions for JevWhat is RLCD?Jev statistics