A new kind of AI model does not write sentences, it picks answers, and three different builders have now shipped one. The real question is whether it becomes the layer that decides when the expensive model gets called at all.
- System One models answer typed questions, picking an option, giving a score or saying yes or no, and return probabilities in one pass
- Jev from TypeSafe started the category on September 15, and the open Laya and Cloudflare's Clef show it spreading beyond one company
- The pitch is speed and cost: Cloudflare reports 38.8 milliseconds for Clef-flash against 524.1 for Jev, though that is a vendor number
- The best fit is the small decisions around an agent, such as routing, classification, verification and guardrails
- It does not replace large language models, which still handle open-ended reasoning, but it could decide when a costly model is needed
- Most benchmarks come from the builders themselves and Laya needs fine-tuning, so test on your own traffic first
Ask a large language model whether an email is spam and it will do something faintly absurd. It spends a few thousand tokens of compute composing a sentence in order to say "yes." Multiply that by every check an AI agent runs while it works, and the sentence-writing starts to look like the most expensive way to flip a switch. A different kind of model has appeared in the last three weeks that does not write the sentence at all, and the idea behind it is now being built by people who have never worked together.
TypeSafe AI started it on September 15 with Jev, which it calls the first System One model. The name comes from Daniel Kahneman's split between fast, intuitive thinking and slow, deliberate thinking. You hand it a block of state and a set of typed questions. It returns structured answers with probabilities attached: a choice from a list, a score on a scale, or a true-or-false verdict. TypeSafe prices it at $0.042 per million input tokens and does not charge for output. Its chief executive Diogo Almeida pitched the whole thing as a question: "Models have been superhuman at chat for years, so where is all the automation?"
What a System One model actually does #
A language model is built to produce language. It reads a prompt and writes a reply one token at a time. That makes it flexible enough to draft a contract, debug a program or reason through an open problem. It also makes it slow and costly for questions with a short, fixed set of answers. A System One model skips the writing. It reads the state and the schema, scores every allowed option at once in a single pass, and hands back numbers your code can branch on. There is no text to parse and nothing to hallucinate in the usual sense, because the answer has to be one of the options you offered.
That makes the territory narrow and quite useful: classification, routing, ranking, verification, picking an agent's next step, checking whether an action matches the task it was given. TypeSafe notes that output tokens on a standard language model typically cost around five times as much as input tokens, which is exactly the part this design removes.
Open Source Is Pulling Away From JEV as Nandakishor Mukkunnoth's LAYA Hits 121 Contributors Nandakishor Mukkunnoth's Laya has 121 contributors and more than 30,000 GitHub stars two weeks in. Why the open source community is pulling away from TypeSafe's Jev, and what has changed on both sides. - open source decision model pulling away from Jev - nandakishor mukkunnoth laya alternative to typesafe jev
Laya and Clef show the idea moving beyond one company #
The more interesting development is who else showed up. Three days after Jev, a developer in Kerala named Nandakishor Mukkunnoth released Laya. It is a 421-million-parameter open model under the Apache 2.0 licence, and it takes the same three question types. We wrote about the developer who says he built Jev first in the first week, and Laya has since passed 30,000 stars on GitHub, with more than a hundred contributors. Its own documentation puts a single answer at about 33 milliseconds on a Tesla T4 graphics card. It is also candid about limits: the base model scores near chance when used as it comes, and it is meant to be fine-tuned on your own examples.
Then on October 1, Cloudflare released Clef. It comes in two sizes, a 27-billion-parameter model for the most accurate calls and a 9-billion-parameter Clef-flash for the latency-critical ones, both built on Qwen backbones and released as open weights under Apache 2.0. Cloudflare says its threat intelligence team uses Clef to categorise website domains. Clef takes images as well as text and handles a 64,000-token window against Jev's 32,000. It also speaks Jev's request format, so code written for one should work with the other. In Cloudflare's own tests, across 43 benchmark runs, Clef-flash answered in a median of 38.8 milliseconds. Clef took 209.3 and Jev 524.1. Cloudflare also says Clef scored highest on seven of ten decision benchmarks.
The large labs have noticed too. OpenAI showed a Decisions API at its DevDay event, built on its GPT-6 Luna model, and its own slide claims 150 milliseconds against 1.6 seconds for a standard Luna call. Amazon released Strands Decider 2B, an open model built on Qwen3.5-2B. It began when distinguished engineer Marc Brooker built a version of his own that briefly topped a Jev leaderboard in its size class. Even the training idea travels: Jev, Laya and Clef each describe rewarding a model for honest probabilities instead of confident guesses.
One thing the lineup makes clear is that "small" is not the defining trait. Laya has 421 million parameters and Amazon's model has 2 billion, but Clef is 27 billion. What they share is the job and the shape of the output, a scored answer in one pass.
Why it matters for agents, developers and businesses #
Agents are where the arithmetic bites. A single task can involve dozens of small judgments. Should it call a tool? Which sub-agent goes next? Does an output look trustworthy, or should an action be blocked? At 1.6 seconds and a full set of generated tokens each, you ration those checks. At tens of milliseconds and no output charge, you can run one on every step. One hackathon demo reported by TechCrunch priced that kind of monitoring at $2.94 with Jev against $372 with a frontier model. It is a single demo's estimate and not a benchmark, but it shows why the cost gap gets attention.
The second benefit is the confidence score. Because every answer comes with a probability, a system can act on the sure ones and escalate the rest, either to a person or to a bigger model. That is the layer idea in practice. A cheap model sits in front, makes the thousands of routine calls, and wakes the expensive one only when a question is genuinely hard or the confidence is low. Open weights add a business angle on top: a team that downloads Laya or Clef can run it on its own hardware, keep its data in-house and avoid depending on one vendor's pricing.
A new layer, not a replacement #
None of this removes the need for large language models. A System One model cannot draft a report, hold a conversation or work through a messy problem, and nobody serious claims it can. The cleaner argument is that these models take on the part of the workload that never needed a paragraph to begin with.
The Developer Who Says He Built Jev First Just Gave Everyone a Free Version Nandakishor Mukkunnoth says he published the idea behind TypeSafe's Jev a year before its launch. His case is only partly right, but his free, open-source answer, Laya, collected more than 23,000 GitHub stars in its first week. - developer who built Jev model before TypeSafe - free open source alternative to TypeSafe Jev
There are reasons for caution. Nearly every benchmark so far was run by the people selling or releasing the model, and each one picks its own tests. Laya's strong results come from a version fine-tuned on 2,000 labelled examples, and it struggles when a choice has dozens of options. Almeida himself has dismissed the first wave of rivals as ML engineers wanting to build a cool architecture more than teams devoted to making intelligence useful. That is an interested view, but it is a fair warning. A fast answer is only worth having if it is right.
So can the future of AI applications be a few powerful reasoning models surrounded by many small deciders? The early evidence says it is plausible. The logic is economic, and the same architecture is turning up in places that have nothing to do with each other. It will be settled by something duller than a leaderboard. Count how many of the calls your product makes today needed a paragraph back, and how many only needed a word.
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.