Jev
TypeSafe's System One model. Typed decisions, no chat. Trained with RLCD. Output tokens free.
TypeSafe Jev · 15 September 2026
Jev builds on X
TypeSafe left stealth on 15 September 2026 with Jev. 43 first-party ships we cataloged, each with the original post.
Diogo Almeida, who worked on InstructGPT at OpenAI, trained Jev with RLCD, Reinforcement Learning for Calibrated Decisions. The model cannot write a sentence. It scores choices. Input is $0.042 per million tokens. Output is free. That is the whole pitch, and people started wiring it into crawlers, code review, and trading bots the same week.
Open weights showed up within days. Laya from Nandakishor M, Kev from Jared Palmer, SemIf from Theodore Lee. Same shape of question, local models, no TypeSafe key. We list first-party posts for those too.
Official note: Introducing System One Models and Jev. Waitlist: typesafe.ai. Announcement: the original post.
Laya
Open Apache 2.0 System One model. ModernBERT, typed choice/score/noul in one pass. Fine-tune notebook on Kaggle T4s. Not Jev.
Kev
Jared Palmer's open Jev-like family on Qwen3.5. 0.8B, 4B, 9B. TypeSafe SDK with a base_url change. Train on one H100.
SemIf
Open decision model that reads option logits from local Qwen-class weights. Browser demo at openjev.com. TypeSafe SDK compatible.
laya-mlx
Laya on Apple MLX. Under 1GB RAM, 7 to 14ms on M3 Max. Snake at 60 decisions a second, no PyTorch, no cloud.
System One Computer Use
Computer use on Jev. Screenshot, decide, click or type, with a log of every step. Open source.
People vs your drawing
Draw something. Jev guesses what it is, then judges the drawing.
laya-apple
Same Laya weights on MLX GPU and Apple Neural Engine at once. P99 1593ms to 80ms on M4 Max.
jev-effort-router
Claude Code plugin. Jev sets effort to low, medium, high, or xhigh per prompt. Fail-open if Jev errors or times out.
vibecheck
Python DSL for Jev. Four functions: check, classify, label, and score. Drop decision models into your code.
Sensibility
Claude Code plugin. Second opinion from Jev: does this diff match the ticket, is this command safe to run.
jev-swap
Scan TypeScript, JavaScript, or Python for LLM calls that are really decisions. Estimates Jev cost and speed. npx jev-swap scan.
Jev Skill Suggestion
Claude Code mod. Jev picks one skill per prompt and injects only that SKILL.md. The rest stay out of the context window.
jevyoumean
CLI wrapper that asks Jev what you meant. git record suggests git commit. npm add suggests npm install. Intent, not edit distance.
openJev Verdict
Open 151M decision model on ModernBERT. JevBench #2 on public tasks after the inference-engine fix. Weights and a WebGPU playground.
APUS-OpenJev
4B, 9B, and 35B-A3B decision family for browser agents. ~26ms median. TypeSafe SDK. Weights on Hugging Face.
SemIf on LangSmith Gateway
LangChain serving SemIf free for a week through the gateway. TypeSafe SDK, change a string. Jev BYOK works too.
ballot
Open System One model in the same class as Jev, Kev, and Laya. State plus typed questions in, calibrated probabilities out.
Jev For Dummies
Self-hosted HTTP GET wrapper for TypeSafe Jev. Paste a URL in the browser and get a typed decision back.
Rumorville
Whisper a rumor to one of 12 residents. Watch Jev decide who gossips and who runs to the cafe.
Play with Jev
Chess against TypeSafe Jev. Open source, playable in the browser.
Cold Read
Paste a cold email. Jev scores four checks in under 500ms. BYOK, runs in the browser.
Jev macOS Loop
Native Mac clicker for Claude Code and Codex. Jev decides, OmniParser and Apple Vision stay on-device. Apple silicon, BYO token.
AnyFilter
Chrome extension that uses Jev to hide anything you describe on X. Ads, spam, or a sentence you type.
Triage Bot
Realtime support triage with Jev routing tickets, then Cerebras drafts the reply. Typed decisions, no text to parse.
quackd
CLI that puts Jev on a LeRobot SO-101 arm. Discrete steps go to Jev, angles stay with the LLM.
RubikJev
Jev solves a 3x3 Rubik's cube in about 0.5s. Playable in the browser, source on GitHub.
LocalJev
Poor man's Jev on local models via omlx. Same API shape as Jev, for people without access yet. From GitHub Next.
Sift
Chrome extension that uses Jev to tag tweets and fold bait, filler, and AI copy.
jevis
Flutter package for natural-language integration tests, scored by Jev.
YouTube Sponsor Detection
Chrome extension that listens to YouTube audio, detects sponsor segments, and skips them in real time. Open source, BYOK.
JevScout
Give it a company homepage. Jev finds Careers, scores roles against your profile. Job crawler rebuilt from 5 minutes to 20 seconds.
jev-review
Simple Jev review bot. Shipped in the first hours of Jev access.
jev-rs
Rust SDK for Jev. Typed async choices, scores, probabilities, and confidence.
jev-harness
Production layer for Jev: policy, confidence gate, shadow mode, recipes, eval CLI.
claude-jev
Wires Jev into Claude Code so every review finding is scored before you read it.
jevscan
Proof-of-concept code vulnerability scanner using Jev. Flags risky code without an LLM writeup.
Ask Jev
Playable Jev demo. A needle that panics in realtime. Possibilities, not text.
jev-pref
Turns AGENTS.md preferences into Jev-shaped checks. An intelligent linter for your agent.
pi-jev-context
Pi extension powered by Jev that hides low-value history from model context without deleting the session.
Jev Trader
Trading bot with Jev. Buy or sell from a price feed, then places orders on Monad via Kuru every 300ms.
Jev, in short
- What is Jev?
- Jev is TypeSafe AI's first public System One model, released 15 September 2026. You send a structured question and get a typed choice, score, or probability. It does not generate prose.
- How is Jev different from ChatGPT or Claude?
- Those models write tokens one at a time. Jev answers in parallel and is not a chatbot. TypeSafe prices input at $0.042 per million tokens and lists output as free. Their own workflow numbers are 20 to 200 times faster than an LLM on decision tasks. We have not re-run those benches.
- Where do I get access?
- Waitlist and docs are on typesafe.ai. The launch note is on the TypeSafe blog. Diogo Almeida posted the announcement on X.
- What counts as a Jev build here?
- A public URL or repo posted on X by the person who made it. Joke wrappers and is-even demos stay off the list.
- Are there open-source alternatives?
- Yes. Laya is Apache 2.0 on ModernBERT. Kev is Jared Palmer's Qwen3.5 family with a TypeSafe-compatible API. SemIf reads option logits from local open models. They answer the same kind of typed question. They are not Jev.
