What is a System One (System 1) model? Decision and extraction models

The short version

A System One model, also written System 1 model, is an AI model that returns typed answers with probabilities instead of generated text. It comes in two classes: decision models such as Jev pick from options you list, and extraction models such as Clairic return typed values read or computed from the input.

TypeSafe AI coined the term. Its own wording, written when every System One model was a decision model, is “System One models make fast, structured decisions for software.”

Most System One models are decision models. A decision model does the job of a classification model, picking a label from a list, and also scores content on a scale and answers yes/no questions, with a probability on every answer.

You hand a decision model some content and a set of typed questions. It hands back answers that match those types, plus a probability for every possible answer. It does not write a reply, produce code, or explain itself. The docs put it plainly: System One models “do not write replies, produce code, or generate explanations of their reasoning.”

Since 19 October 2026 there is a second class. Cloudflare calls its new model Clairic “a new class of System One-style models: extraction models”. An extraction model returns typed values instead of decisions: an amount, a date, a span of the input, a forecast. TypeSafe coined System One and had not commented on the new class as of 19 October 2026. The rest of this guide is mostly about decision models, because every System One model before Clairic was one.

Where the name comes from

Daniel Kahneman’s Thinking, Fast and Slow splits human thought in two. System 1 is fast and intuitive. You see a face and know it is angry before you could say why. System 2 is slow and deliberate. You multiply 17 by 24 on paper.

TypeSafe’s launch post borrows that split, contrasting “fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning.” The claim is about the shape of the task, not about what happens inside the model. Sorting a support ticket into one of five queues is a System 1 task for a person. Writing the reply is not.

Kahneman was describing people. Nobody has shown that a model trained on these tasks works the way human intuition works, and the Thinking, Fast and Slow entry keeps that distinction clear.

The industry now calls Jev’s class decision models, and What is a decision model? covers that name.

What counts as a System One task

For a decision model, a System One task is a judgment with a known set of possible answers, needed quickly, where nobody has to see the reasoning. Sorting a ticket into a queue is one. So is deciding whether a passage is relevant, rating urgency on a scale, or screening text for a policy breach.

TypeSafe never defines the phrase on a page of its own. It uses it descriptively, saying on the jaggedness page that “jev-1.13 does the best on System One tasks.” The same page gives the only named contrast in the docs, a warning to avoid “System Two tasks: more layers of indirection.”

That leaves the failures to mark the boundary. Counting, arithmetic and date comparison fall outside it, as does anything that chains two facts together or ends in a paragraph. The eighteen patterns in the use cases section are all the other kind: routing, filtering, ranking, gating and scoring.

Extraction moves part of that boundary. Cloudflare built Clairic for questions whose answers cannot be listed in advance, such as an invoice total or a delivery date, and says it handles one- and two-step calculations reliably. Longer chains still fall outside.

Two classes: decision and extraction

Cloudflare’s launch post for Clairic sets the two classes beside LLMs in one table. A condensed version:

The post sums it up: “Extraction models turn raw input into typed values, decision models turn values into decisions, and LLMs turn decisions into words.” This is Cloudflare’s classification. The extraction model guide covers the new class in detail.

What comes back

For a decision model, three things, depending on the question type. A typed answer: the option you defined, a number on the scale you defined, or a probability. A full probability distribution over the options you gave. And, for two of the three question types, a confidence value between 0 and 1.

Confidence is not a second opinion from the model. TypeSafe describes it as “a statistic computed from the probability distribution the answer already gives you.” A tight distribution means high confidence. A flat one means the model is guessing. Your code decides what to do about it, which is the whole point of confidence gating.

The three primitives

Every question you ask a decision model is one of three shapes.

Choice picks one option from a list you define and returns the winner, a probability for each option, and a confidence value. It takes up to 255 options.

Score places the content on an ordered scale you define. It needs at least 2 levels and takes at most 10. The returned score is the probability-weighted mean of the level numbers, so it can land between levels: (0 x 0.0) + (1 x 0.70) + (2 x 0.30) gives 1.30.

Noul asks a yes/no question and returns a single number from 0 to 1, the probability that the answer is yes. It returns no separate confidence value, because the probability already is one. The docs do not explain where the name came from.

The Choice, Score and Noul guide works through a full example of each.

Why they cannot hallucinate free text

A System One model never generates prose, so there is no sentence in which it can invent a citation, a policy number, or a customer name. A decision model’s output space is the set of options you wrote down; an extraction model’s is the type you declared, or null. That is a real property.

It is also a narrow one. The model can still be wrong. It can pick the wrong option with high confidence, or return a probability that does not match how often it is actually right. TypeSafe publishes a per-version page of known weaknesses it calls jaggedness, and the jaggedness guide lists the ones that bite hardest: Jev 1.13 reads questions literally, is weak at arithmetic and counting, “reads dates as text, not as ordered quantities,” and by default treats everything in the input as data rather than as potentially hostile. Losing free-text hallucination does not buy you correctness.

Cloudflare makes the same point for extraction. Clairic’s answer is “a value of the type you asked for, or null”, and “the architecture makes every answer well formed, not necessarily correct.”

How they differ from LLMs and reasoning models

An LLM predicts the next token, then the next, conditioned on everything before it. Ask it to classify something and it still runs a generation loop to produce the word “billing.” A reasoning model does more of that loop, spending extra tokens on intermediate steps before it answers.

A System One model skips the loop. TypeSafe says Jev “generates all outputs in a single query” rather than one token at a time. Its architecture is unpublished, so how that is done is a guess for now. See how Jev is built for what is known and what is not.

The practical differences: an LLM returns a string you have to parse and validate, a System One model returns a typed value you can branch on. An LLM gives you a token log-probability if you dig for one, a System One model gives you a calibrated distribution as the primary output. The side-by-side comparison covers cost, latency and failure modes.

Where they fit in a stack

Inside the code, not around it. TypeSafe’s own framing is that the model “does not generate code or choose its own next action.” You keep the control flow, the thresholds and every side effect. The model answers narrow questions about state you assembled.

A typical shape: a System One model decides which queue a ticket belongs in and how urgent it is, your code applies the routing rules, and an LLM writes the customer-facing reply once the decision is made. The two model types are not competing for the same slot. An extraction model can sit in front: it pulls the values out of the raw input, and the decision model decides on them. Confidence-gated actions and LLM guardrails are two of the places this pattern shows up most often.

Which System One models exist

Jev came first. TypeSafe AI coined the category name and announced Jev on 15 September 2026 alongside a $40M seed round led by DCVC. It is a closed managed API and the weights are not released. The Jev model page has the specs, and Jev explained is the plain-English version.

Others followed over the next five weeks:

• Kev, an open-source family from Jared Palmer, from 0.8B to 27B on Qwen3.5 and Qwen3.8 bases, announced 17 September 2026. It serves the same /v1/systemone request shape as TypeSafe’s API. Since 25 September 2026 you can also call Kev-4B on OpenRouter at $0.042 per million input tokens.
• Laya, open weights from Convai Innovations on ModernBERT and mmBERT encoders, announced 18 September 2026.
• Surogate Rune 26B-A4B, open weights from Surogate under Apache 2.0, announced 21 September 2026. A full fine-tune of Gemma 4 26B-A4B-it that answers Choice, Score and Noul questions and reads images.
• Solar Decide, Upstage’s model on Solar Mini 4 with a 512K-token context, in beta on Upstage’s API since 22 September 2026. It answers Choice, Score and Noul questions on the same /v1/systemone request schema as Jev.
• Decider 1, meraGPT’s hosted decision model, announced 22 September 2026. It uses the same /v1/systemone request schema at $0.03 per million input tokens, with up to 4,096 tokens per request.
• Tev1, Together AI’s Jev-like classifier on Qwen3.5-4B, released 23 September 2026 on Together’s serverless API. It picks one option from a list of 2 to 24.
• CLM, open weights from a research group led by Jacky Kwok, released 23 September 2026. It scores a state against candidate actions with a contrastive objective.
• GLiNER2.5-Decide, open weights from Fastino Labs, released 24 September 2026. A 340M-parameter encoder that answers typed classification questions and extracts spans, relations and structured records in one pass.
• Span-01, Respan’s classifier for agent traces, announced 24 September 2026. It returns the probability that each behavior you define is present, absent or not observable, through Respan’s own API.
• Drex, Nace.AI’s decision model family, launched 24 September 2026. It uses Jev’s /v1/systemone request format at $0.06 per million input tokens for Drex Max and Drex 2.0. Drex Max reads images, and Drex 2.0 has open weights under a licence that limits commercial use.
• Kai, Hanzo AI’s decision model, launched 28 September 2026. It answers Choice, Score and Noul questions at $0.021 per million input tokens through Hanzo’s /v1/decisions endpoint, with closed weights.
• d1, Liquid AI’s first decision model, announced 29 September 2026. It answers Choice, Score and Noul questions on Liquid’s API as d1:free or the paid d1, and on OpenRouter at $0.04 per million input tokens. On 7 October 2026 Liquid added two open-weight versions, d1-3B and d1-omni-600M.
• Strands-Decider 2B, open weights from Strands Labs under Apache 2.0, announced 29 September 2026. A 1.9B model on Qwen3.5-2B-Base, aimed at the small decisions inside an agent, served over /v1/systemone on your own machine.
• The OpenAI Decisions API, announced 29 September 2026 and powered by GPT-6 Luna. It answers typed questions about text or images with probabilities, and has been in public beta since 6 October 2026 at $0.10 per million input tokens.
• Mercury Decide, from Inception, launched 30 September 2026. It answers Choice, Score and Noul questions at $0.02 per million input tokens on Inception’s API and on OpenRouter, which also has a free tier.
• GLiDE, Fastino’s decision model, launched 30 September 2026. It answers Choice, Score and Noul questions on Fastino’s /v1/systemone endpoint at $0.04 per million input tokens. Version 1.2, released 22 October 2026, answers in one forward pass, reads images, and has weights on Hugging Face under a licence that limits production use by large companies.
• Clef, Cloudflare’s own decision model in 27B and 9B sizes, launched 1 October 2026 on Workers AI. It uses Jev’s request format, reads images, and has open Apache 2.0 weights.
• pplx-decider, Perplexity’s decision model, launched 1 October 2026 through its Decisions API, now in a 27B and a 4B Flash size. It reads images, has open Apache 2.0 weights, and uses Perplexity’s own /v1/decisions request format instead of Jev’s.
• H2O-Lightning-4B, open weights from H2O.ai under Apache 2.0, announced 2 October 2026. A fine-tune of Qwen3.5-4B that answers Choice, Score and Noul questions, reads images, and serves Jev’s /v1/systemone request shape. H2O has not published a hosted API.
• Quyet-1.0-Large, open weights from Chinh Nguyen under Apache 2.0, announced 3 October 2026. A 31.3B LoRA fine-tune of Gemma-4-31B-it, the largest of five Quyet 1.0 sizes, with answers in the /v1/systemone shape.
• Torchcast Decision 27B, from Torchcast AI, announced 3 October 2026. It uses Jev’s /v1/systemone request format, and its weights are CC BY-NC 4.0, so non-commercial use only.
• Celeris-1 Decision, from Celeris, launched 8 October 2026. It reads images, uses Jev’s /v1/systemone request format, costs $0.04 per million input tokens and can add a one-sentence reason to each answer.
• Microsoft-Decision-1, Microsoft’s decision model, announced 9 October 2026 on Microsoft Foundry. Post-trained from Qwen3.5-9B, it returns a calibrated probability for each fixed option at $0.042 per million input tokens, with closed weights.
• Clef-omni, Cloudflare’s third Clef model, launched 9 October 2026 on Workers AI. It reads audio and video as well as text and images, costs $0.15 per million input tokens, and has open Apache 2.0 weights.
• SubQ Snap, Subquadratic’s decision model, first released 12 October 2026 and in public preview. It reads up to 384K tokens of text and images per request on its own /v1/systemone endpoint at $0.04 per million input tokens, with closed weights.
• d1-ultra, Liquid AI’s larger hosted decision model, released 15 October 2026. It answers Choice, Score and Noul questions about text and images at $0.042 per million input tokens, the same as Jev, on Liquid’s API and OpenRouter.
• Span-02, Respan’s general decision model, released 15 October 2026 and a different model from Span-01. It answers Choice, Score, Noul and number questions about text and images, can think briefly before answering within a deadline you set, costs $0.01 per million input tokens, and has open weights on Gemma 4 26B-A4B.
• Jeffan-31b-1710, open weights from the Hugging Face account of Trong Vu, published 19 October 2026. A full fine-tune of Gemma-4-31B-it under Apache 2.0, second on the Jev Decision Index 0.4.1.
• H2O-Lightning-31B, H2O.ai’s larger decision model on Gemma-4-31B-it, published 20 October 2026. It reads images and serves the /v1/systemone request shape, and its weights are under a non-commercial licence.
• mammoth-1 and elephant-1, 27B and 9B decision models from Baseten’s research team Base Labs, released 21 October 2026. Both serve Jev’s /v1/systemone request on Baseten’s Model APIs, free for their first week, and have MIT weights on Hugging Face.

All of those are decision models. The first extraction model is Clairic, from Cloudflare, in preview on Workers AI since 19 October 2026 at $0.10 per million input tokens. It reads text and images and returns typed values with probabilities. Its API takes named fields, not Jev’s questions.

None of them uses TypeSafe’s training method, RLCD, because TypeSafe has not published a paper or dataset for it. For a verdict on the hosted, open-source and local options, see Jev alternatives. The models directory lists everything the site tracks, and the examples directory tracks what people have built.

System 1 AI, System 1 LLM and System 1 decision engines

TypeSafe’s term is System One model. Used in that sense, “System 1 AI” and “System 1 decision engine” describe the same thing: a model that returns a typed answer your code acts on, a decision for a decision model or a value for an extraction model. “System 1 LLM” is the misleading one, because a System One model writes no text and so is not an LLM. AI researchers also use “System 1” more loosely for a language model that answers directly, without writing out reasoning steps first, as the 2024 paper Distilling System 2 into System 1 does. That kind of System 1 LLM still generates text, which makes it a different thing from a System One model.

FAQ

Is a System One model an LLM?

Not in the usual sense. It takes text input and makes judgments about it or reads values out of it, but it does not generate text. TypeSafe has not published Jev’s architecture, so whether a transformer language model sits underneath is unknown. Treat “System One model” as a description of the interface and the training target rather than the internals.

Can a System One model hallucinate?

It cannot invent free text, because it never writes any. A decision model’s answers are always members of the option set you defined; an extraction model returns a value of the declared type, or null. Either can still be wrong, or return a probability that does not match its real accuracy. TypeSafe documents nine known weakness areas for Jev 1.13, including arithmetic and counting.

Do I still need an LLM?

Usually yes. A System One model decides or extracts; it cannot draft an email, write code, or summarise a document. The common pattern is a System One model making the routing and gating decisions inside your code, with an LLM handling any text a human will actually read.

What does calibrated mean here?

A model is calibrated when its stated probabilities match reality: of all the answers it gives at 80% confidence, roughly 80% should turn out correct. That property is what lets your code set a numeric threshold and trust it. TypeSafe has not published a calibration error figure for Jev.

Is a System One model the same as a classification model?

A decision model does a classifier’s job, and then some. A classification model sorts input into categories, which is what a Choice question does. A decision model also scores against ordered levels, answers yes/no questions with a probability, takes its labels per request rather than at training time, and is trained so those probabilities match real outcome rates. Whether that earns a new name is a live argument.

Who else makes a System One model?

Hosted decision models come from Together AI, Upstage, Respan, Baseten, Liquid AI (d1 and d1-ultra), meraGPT, Inception, Fastino, Cloudflare (Clef and Clef-omni), Perplexity, Hanzo AI, Nace.AI, Celeris, Subquadratic, Microsoft and OpenAI. Kev, Laya, Surogate Rune, CLM, GLiNER2.5-Decide, Strands-Decider, H2O-Lightning-4B, Quyet-1.0-Large and Jeffan-31b-1710 are open source. Torchcast Decision 27B and H2O-Lightning-31B are non-commercial. Cloudflare’s Clairic, in preview since 19 October, is the first extraction model. The models directory lists each one.