LLMs are great at chat. Jev from TypeSafe AI is built for something else: fast decisions your code can use directly.
It comes from Diogo Almeida, who worked on the research behind ChatGPT at OpenAI. After two years in stealth, his team released Jev as their first System One Model, trained with a new method called RLCD: Reinforcement Learning for Calibrated Decisions.
This post is a simple breakdown: what Jev is, how it differs from LLMs, and where to access it.
What is Jev in simple terms
Send in messy state plus structured questions. Get back typed answers with probabilities.
No chat text. No essay. Just values like approved: true, confidence: 0.92 that software can branch on.
TypeSafe calls it a frontier intelligence function call: unstructured state in, typed probabilistic decisions out.
How Jev differs from LLMs
| Area | LLMs | Jev |
|---|---|---|
| Optimized for | Human preference with RLHF, or verifiable rewards with RLVR | Calibrated decisions with RLCD |
| Input | Sequential messages | Structured program state |
| Output | Strings that need parsing, can hallucinate | Type safe values, zero type errors by design |
| Sampling | Sequential, one token at a time | Parallel, all outputs in one query |
| Confidence | Often overconfident and inconsistent | Always returns calibrated confidence |
| Speed | 3 to 329 seconds end to end for frontier models | 70ms to 500ms, claimed 20-200x faster |
| Cost | $0.20 to $10 per million input tokens, outputs about 5x more | $0.042 per million input tokens, outputs free |
The tradeoff is clear: Jev cannot generate text. If you need writing, code, or chat, use an LLM. If you need classify, route, score, extract, or verify at scale, Jev fits.
Demos that show the idea
- Side by side demo: same workflow run on Jev and GPT-5.6 Terra. Jev outputs all probabilities in parallel instead of autoregressively.
- Doom bot: real time control at about 10 queries per second for roughly $7 per hour. Built for automation, not just chat.
- Wikiracing: start on one Wikipedia page and reach another using only links. Great test for high cardinality choices up to 255 options per step, with no hallucinations.
Full evals, queries, and disagreements are published at evals.typesafe.ai.
Where to access Jev and how
Jev is in early access today:
- Main site and waitlist: typesafe.ai
- Launch post: Introducing System One Models and Jev
- Docs: docs.typesafe.ai
- Playground: console.typesafe.ai/playground
- LLM comparison adapter: system-one-adapter-python on GitHub
- Updates: @typesafeai and @CompleteSkeptic
How to try it:
- Join the waitlist on typesafe.ai
- Read the docs for the query format:
stateplusquestionswith a predefined schema - Test in the playground, then call the API from code and threshold on confidence: act when high, escalate for review when low
Closing thought
LLMs are talkers. Jev is a decider.
Use LLMs for human in the loop tasks like chatbots, copilots, and coding agents. Use Jev for smart if statements buried in workflows, map reducing over big data, real time apps, and verifying LLM outputs.
Sources: the TypeSafe launch blog, typesafe.ai homepage, and the original X announcement plus discussion.