Jev AI guide · updated September 18, 2026
What is Jev?
Jev is the flagship model from TypeSafe and the first System One model. You give Jev a piece of text and a set of narrow, typed questions. Jev answers every question at once with a probability, a choice or a score, and says how confident it is. It never writes prose.
System One: fast, typed judgments
The name comes from Daniel Kahneman's Thinking, Fast and Slow. System 1 thinking is fast and intuitive; System 2 is slow and deliberate. Jev targets the first kind: quick, focused judgments that software can branch on.
A request has three parts. The state is the text or JSON Jev reads. The model is jev-latest, which currently resolves to jev-1.13.0. The questions are a map of typed primitives:
- Noul — a yes/no question. Jev returns the probability of yes, from 0 to 1.
- Choice — pick one option from a set you define. Jev returns the winner, every option's probability and a confidence.
- Score — rate against ordered levels you write. Jev returns a probability-weighted score that can land between levels.
Jev reads the state once and evaluates every question in parallel. You can see this in the Jev AI demos: four questions about a support ticket come back in one response.
How Jev is trained: RLCD
TypeSafe describes three ways to post-train a pretrained language model.
- RLHF (reinforcement learning from human feedback) trains chatbots to give answers people prefer. TypeSafe points out that it can reward confident-sounding mistakes and narrow the range of outputs.
- RLVR (reinforcement learning with verifiable rewards) produced reasoning models. They are strong at maths but slower and more expensive.
- RLCD (reinforcement learning for calibrated decisions) is Jev's path. The model returns decisions and probabilities, and higher probability should mean a higher chance of being right.
Calibration is a property of many predictions, not a promise about one. Across many answers where Jev says 0.8, about 80% should be correct. That is what makes the confidence field useful: act automatically when it is high, ask for review when it is middling, and route to a person when it is low.
Jev model facts
| Current version | jev-1.13.0 (aliases jev-latest and jev-preview) |
| Input | Text only: a string, a JSON object or an array. No images, audio or video. |
| Context | 64k tokens per request; 32k for the state plus the longest single question |
| Official price | $0.042 per million input tokens on TypeSafe; output tokens are free |
| Official rate limits | 250,000 tokens per second / 1,200 requests per minute, adjusted as capacity changes |
| Language | Best in English; other languages work with lower accuracy |
| Customer data | TypeSafe states that Jev is not trained on customer requests and is not fine-tuned per account |
Facts from TypeSafe's public documentation as of September 18, 2026. Jev AI pricing is on the pricing page.
Where Jev is strong, and where it is not
TypeSafe publishes a candid list of Jev 1.13's weak spots. Designing around them is most of the craft.
- Good at: common-sense judgments about text: routing, moderation, intent, sentiment, relevance, policy checks, citation support and guardrails.
- Literal reading: Jev answers the question you wrote. Put boundary cases in the criteria.
- Numbers, counting and dates: keep arithmetic and date comparison in code. Use Jev to extract or classify the pieces.
- Long, noisy state: accuracy falls when the state is full of unrelated text. Filter first.
- Adversarial text: injected instructions inside the state can move answers. Test edge cases before launch.
- Generation: Jev does not write. Pair it with a generative model when you need text.
How to use Jev
- 01
Try it without code
Open the Jev AI playground. Load a demo, edit the questions and run it on your own text.
- 02
Copy the request
“Get API code” turns your playground setup into cURL, JavaScript or Python.
- 03
Call the Jev API
Create a key on the Jev API page and send the same JSON to
/api/v1/systemone. You can also use TypeSafe's own API and SDKs directly.
More questions about Jev
Who makes Jev?
Jev is built by TypeSafe. TypeSafe's documentation notes that cofounder Diogo Almeida co-invented RLHF, the technique used to train InstructGPT and ChatGPT.
Is Jev an LLM?
Jev understands natural language like an LLM, but it does not generate text. It is a System One model that returns typed decisions and probabilities.
What does RLCD mean?
RLCD stands for reinforcement learning for calibrated decisions, TypeSafe's post-training method. It rewards correct decisions with honest probabilities rather than answers people prefer to read.
Is there a Jev benchmark?
TypeSafe publishes task cookbooks rather than a single leaderboard score. In its re-ranking cookbook, one Jev question per candidate raised top-1 accuracy on a legal retrieval set from 5% to 18%. Batching 13 questions into one call was 12.2x cheaper and 10x faster than separate calls. Test Jev on your own data before you rely on it.
Can I use Jev on OpenRouter or Vercel?
Jev has its own API shape (state plus typed questions), not the chat-completions format, so general chat routers do not map onto it directly. Use the TypeSafe API, its Python and JavaScript SDKs, or the Jev AI API on this site.
Jev AI is an independent service and is not affiliated with or endorsed by TypeSafe. Jev answers are probabilistic; review important decisions before acting on them.