Choice
Pick one option from a set you define — up to hundreds of routes, labels, or moves.
Unofficial notes on Jev
imjev
It doesn't chat. It plays.
Typed decisions fast enough for Doom, Mario, and software that can't wait on a chat model.
Not chat.
RLHF made models great at talking to people. That same pressure makes them shaky when software needs a yes, a route, or a move — overconfident prose, mode drop, humans glued to the loop.
Decisions.
Jev takes structured state plus typed questions and returns choices, scores, and probabilities your code can act on. No essay. No JSON surgery. More like a function than a chatbot.
What is Jev AI?
A System One model: built for machine-speed judgment inside workflows, not for conversation.
Score
Place the state on a short rubric (for example 1–5 urgency or quality).
Noul
Yes/no as a calibrated probability — act when high, escalate when not.
Latency
~70–500ms
decision pass, not token stream
Input price
$42 / B tokens
per TypeSafe launch materials
Output
Typed + confidence
software can threshold & branch
Figures and product claims come from typesafe.ai and public write-ups. Verify against official docs before you ship.
Why Jev is trending
TypeSafe came out of stealth in September 2026 with Jev — a System One model built for machine decisions, not chat. Within days, X was full of real-time demos.
The launch story
TypeSafe AI announced System One models and Jev as a frontier model for automation: structured state in, typed decisions with calibrated confidence out — often cited around 70–500ms.
TypeSafe: Introducing System One & JevPress: Doom, not dialogue
Coverage framed Jev as a model for machines — probabilistic typed outputs, not assistant chat — with launch demos including Doom on structured game state.
The Register on Jev playing DoomBuilders on X
Mario in real time, blitz chess vs frontier LLMs, Pac-Man with Astra planning and Jev executing, bulk classifiers for cents — the feed proved latency and cost unlock control loops.
Live Jev posts on XDeep dives
Independent write-ups unpack Choice / Score / Noul, pricing, and how Jev differs from JSON mode — useful if you're evaluating System One workflows.
Flavio Copes: deep dive into JevJev demos lighting up X
The feed didn't argue about chat quality. It showed control loops — games, classifiers, harnesses — running on cheap decisions.
-
01Doom on structured state
Feed a text game state, ask for goals and inputs, run at roughly ten queries a second. The launch clip that made “decision model” feel physical.
The Register coverage -
02Super Mario in real time
Fast inference plus structured outputs steering Mario live — the clip that proved the latency story outside a lab graph.
Find on X -
03Blitz chess vs frontier LLMs
Every move is one API call. Jev V13 in 5+0 against Fable 5.1 and GPT-6 Astra — a stress test of decision speed, not elo bragging rights.
Find on X -
04Pac-Man: Astra plans, Jev executes
A reasoning model sets strategy; Jev carries moves in milliseconds. Built in minutes — the hybrid pattern people are excited about.
Find on X -
05Tetris rotate / move / drop
Repeated micro-choices on board state. Same primitive as ticket routing — just with gravity.
Deep dive write-up -
06Bulk paper classification
Early tests labeled ~1,000 AI papers across dozens of topics for cents, with sub-second median latency — map-reduce shaped work.
Flavio Copes notes -
07Wikiracing without fake links
Choose among hundreds of real outbound links per step. Typed choice means the model can't invent a URL that isn't there.
Mentioned in early experiments -
08Driving / FSD-style control loops
Structured observations in; accelerate / brake / turn out. Builders rebuilt toy self-driving stacks in under an hour to show the loop.
Find on X
Why it blew up
Not because another chat model scored higher on a leaderboard — because the unit of work got cheap enough to put inside a frame.
Latency is a feature
When a decision returns in under half a second, you can ask every tick. Games made that visceral; production systems get the same unlock for routing and verification.
Cost unlocks volume
At tens of dollars per billion input tokens, labeling and filtering stop being “call the expensive model once” and become “spray intelligence across the corpus.”
Cascade, don't replace
The interesting pattern on X: big models think, Jev acts. Astra plans Pac-Man; Jev mashes the milliseconds. Same idea for support desks and agent harnesses.
FAQ
Short answers. Official detail lives on TypeSafe.
Is imjev.com official?
No. This is an unofficial community page. Product, pricing, and API access are owned by TypeSafe AI.
Can Jev chat or write code?
No. It returns typed decisions (choice / score / noul) with confidence. Pair it with a generative model when you need prose or code.
How is this different from JSON mode?
Structured output still starts as text generation. Jev is built to emit decisions and calibrated probabilities for software to consume — TypeSafe's framing is System One, not “chat with a schema.”
How do I try it?
Start at typesafe.ai and their console / early access flow. This site does not proxy the API.
Can it still be wrong?
Yes. Confidence is not omniscience. The point is your software can set thresholds: act, retry, or escalate to a human or a larger model.
Go play with the real thing.
When you're done browsing the feed artifacts, the model lives on TypeSafe.