Ask questions about any text and get typed, calibrated answers — a choice, a score, or a yes/no with a probability — in about half a second. Decisions your software can act on, not prose to read.
Pick a case or write your own, edit the questions, and hit run. Every answer below is a real API call.
Jev is TypeSafe AI's first System One model. Jev AI lets you try it in the browser and call it by API.
Numbers and labels your code can branch on — never a paragraph to parse.
Every answer carries an honest confidence, trained with RLCD.
Ask a choice, a score and a yes/no together — all answered in parallel.
A single typed POST — TypeSafe-compatible, drop-in from any language.
Declare the shape you want and get exactly that back — typed, with the probabilities behind it.
A probability between 0 and 1 that the answer is yes. For gates, checks and guardrails.
Returns one of your keys with a probability for every option. For routing and classification.
A number on an ordered scale you define, with a distribution. For priority, risk and quality.
Start from a ready example — support triage, moderation, routing — or write your own.
Declare the answers you want: a choice, a score, a yes/no. Ask several at once.
One real call returns typed, calibrated answers in ~70–500ms. Free in the browser.
Grab a jv_live_ key and it's a single POST — about 30 lines in TypeScript.
Weighing Jev AI against an LLM, the open/local route, or open-source builds? Each has a full breakdown.
Typed, calibrated decisions vs free text — when to use each, and how to run both together.
Read the comparison →Hosted & calibrated vs the open, local route (Laya-MLX on Apple Silicon). Cost, speed, setup.
Read the comparison →The model is closed-weight & hosted; the tooling (jev-mcp) is open. Open alternatives compared.
Read the comparison →Try the full model free, no card — right here in the browser.
Output is free. About $0.001 per decision at ~$0.42 / M input tokens.
Buy credits or bill by tokens — same hosted model either way.
Jev AI is a hosted decision model from TypeSafe AI. Instead of writing text like an LLM, it returns a typed, calibrated answer — a choice, a score, or a yes/no with a probability — in about 70–500ms, so software can act on it directly.
Yes. The playground on this page runs the real model free, with no signup — you get welcome credits to start. To build with it you buy a jv_live_ key; calls cost about $0.001 each, billed on input tokens with output free.
No. LLMs generate free text for people to read; Jev AI returns a typed value for programs to act on — far faster and cheaper. They're often used together: the LLM writes, Jev decides.
TypeSafe AI. Jev is their "System One" model — fast, instinctive judgement — served over a hosted API.
The model itself is closed-weight and hosted, so there are no weights to download. The surrounding tooling — like the jev-mcp server on GitHub — is open. Open-weight, Jev-like alternatives such as Laya exist for local use.
About 70–500ms per decision — roughly 40–200× faster than a frontier LLM, because it returns a small typed answer instead of generating paragraphs.
Yes. Pass several questions in one call and each comes back independently typed — classify, score and gate the same input in a single request.
Routing, classification, scoring, moderation, and safety gates inside software and AI agents — anywhere you need a fast, machine-readable decision instead of prose.
Run a real, typed, calibrated decision in the browser — free, no signup — then grab a key to build with it.