JEV, explained for people who already understand LLMs

Most people are hearing JEV, JEV, JEV again and again. Far fewer can tell you what it is, where it sits, or what it is bad at. This is the short version, written for someone who already has a working idea of what an LLM does.

Start with what it isn't#

JEV is not a chatbot. It is not an LLM. It is a transformer, but it does not generate text, not a word of it. You cannot ask it to write code, draft an email, or explain its reasoning. That capability was deliberately removed, and removing it is the entire point.

What it actually does#

You give JEV two things: some unstructured state, and a set of questions whose possible answers you have defined in advance. It gives you back typed answers with a probability attached to each one.

Say a message lands in your queue: "The deploy failed twice and customers are seeing 500s. Can someone look now?" You ask JEV whether it is urgent and which team owns it. Roughly 300 milliseconds later you get urgent: yes (0.94) and category: infra (0.81).

There are three question shapes:

• Choice picks one of up to 255 options.
• Score rates something on ordered levels.
• Noul is a yes/no that returns the probability the statement is true.

You can ask many at once and they are evaluated in parallel, so the tenth question costs you almost no extra time.

Why it is fast, and why that is structural#

An LLM produces one token, feeds it back into itself, and produces the next. That loop is why latency scales with how much it says, and why even a one-word answer wrapped in JSON has to be spelled out character by character.

JEV has no loop. Every answer comes out of a single forward pass, all at once. End-to-end latency is 70 to 500 milliseconds. Input costs $0.042 per million tokens and output is free, because there is no output to meter in the usual sense. TypeSafe puts the headline numbers at up to 193x faster and 445x cheaper than frontier LLMs on their own production-workflow benchmarks.

They call the category a "System One" model, borrowing Kahneman's split between fast intuition and slow deliberation. Everything the industry has shipped so far is System 2: thoughtful, verbose, expensive. The bet is that most decisions inside real software never needed System 2. They needed an instant yes or no, and until now the cheapest way to get one was to make an essay-writing machine produce a one-word essay.

Where it fits in your stack#

This is the part most of the hype skips. JEV does not replace your LLM. It sits in front of it.

• Routing. JEV reads the incoming request and decides whether it needs a cheap model or an expensive one.
• Guardrails. JEV classifies a tool call as risky or safe before the agent is allowed to execute it.
• Volume classification. Ticket triage, email sorting, sentiment over millions of rows: the work that is too expensive to send to a frontier model and too fuzzy for a regex.
• Real-time loops. Browser agents, trading agents, games. Anywhere a second is too slow.

The pattern that makes it work is the confidence number. JEV is trained with something TypeSafe calls Reinforcement Learning for Calibrated Decisions, the goal being that a 0.7 actually means 0.7. So you set a threshold: high confidence resolves immediately, low confidence escalates to the big model. Cheap layer in front, expensive layer behind, and a clean handoff between them.

The honest caveats#

Beyond that, it never explains itself, which rules it out of anywhere you need an auditable reason. It will not design your schema for you, and a badly-shaped question produces a confidently useless answer. And the benchmark numbers are TypeSafe's own, measured against model-derived references rather than independent ground truth, so treat them as workload-specific claims rather than universal multipliers.

A fair number of engineers look at all this and conclude it is a zero-shot classifier with unusually good marketing. Sean Goedecke has pointed out that fast structured output is already available by prefilling an existing LLM.

So what is it worth#

Probably this. Even if the skeptics are right about the architecture, they are wrong about the framing. The last three years taught everyone to reach for a text generator whenever a decision looked hard, and to pay text-generator prices for answers worth one bit. JEV is the correction. Whether or not this specific model wins, the shape of it, intelligence as a cheap, typed, sub-second function call, is going to stay.

Built by Diogo Almeida (ex-OpenAI, ChatGPT and RLHF) with Erik Gafni and Sasha Sheng, at TypeSafe AI. Early access opened mid-September 2026.