agi.jetzt
DEEN

Model landscape

System One & Decision Models

A young class of models returns no text but a choice, a score or a probability that software processes directly. TypeSafe AI calls them "System One Models"; agi.jetzt uses the broader term decision models.

Last checked: 10 Oct 2026

What this class is

A decision model receives a state — a message, a record, the current state of an application — and questions whose possible answers are fixed in advance. It returns one of those answers together with probabilities. It writes no reply, no code and no explanation.

The answer is made for software: a program reads it, compares it with a threshold and acts, routes the case onward or hands it to a person.

What is it optimized for?
For one narrow thing: a fast answer from a fixed set of possibilities, together with a probability that software can process.
How does it differ from a generalist?
A generalist writes an open answer for a reader. A decision model picks among answers that were defined before the call, for a program. What differs is the interface, the form of the output, the intended use, and how much latency and calibration matter.
What does it output?
A choice among options, a yes/no probability, a score on a scale — each with probabilities. No text, no code, no explanation.
Where is it useful?
Wherever software has to decide many times and quickly: routing requests, checking the output of other models, approving or stopping a step of an agent, sorting large amounts of data.

Two terms, not the same thing

"System One Models" is a name TypeSafe AI introduced with its model Jev in September 2026. TypeSafe’s documentation defines them as "a class of AI models built to make fast, structured decisions that software can use directly" and takes the name from Daniel Kahneman’s "Thinking, Fast and Slow": System 1 is fast and intuitive, System 2 slower and more deliberate.

It is not an established industry taxonomy. Other providers describe the same kind of model as a "decision model" — Cloudflare, Liquid AI and H2O.ai use that term; Strands Labs writes "decision models or system one models"; Microsoft speaks of a "decision-scoring model" and OpenAI of a "Decisions API".

agi.jetzt therefore uses decision models as the name of the class and "System One Models" for what TypeSafe calls so. The class overlaps with older ones: Microsoft’s own model catalog files Microsoft-Decision-1 under text classification.

Sources: TypeSafe docs: System One, Cloudflare: Introducing Clef, Liquid AI docs: Decision models, Strands Agents: Introducing Strands Decider, Microsoft Foundry model catalog: Microsoft-Decision-1

What is different from a general-purpose model

The difference lies in what the model returns and which interface it is built for.

General-purpose generative model
  1. Context
  2. General-purpose model
  3. Text, code or media

The output is open: anything that can be written. Software has to parse and check it before it can act on it.

Decision model
  1. State + typed question
  2. Specialized decision model
  3. Choice, score or probability
  4. Software action

The possible outputs are fixed before the call. The answer is one of them, with probabilities, and software can act on it directly.

Not: probabilistic versus not probabilisticGeneral-purpose language models are probabilistic too: they compute a probability for every possible next token. The distinction lies in four things: the interface the model is optimized for, the form of its output, its intended use, and how much latency and calibration matter — a generated text for a reader on one side, a probability over fixed answers for a program on the other. OpenAI’s Decisions API offers such answers and names gpt-6-luna as the model behind it.

Eight entries, in four groups

Jev is not the whole class. The groups say what kind of thing each entry is; nothing here is an order of merit.

TypeSafe’s "System One Models"

The name is TypeSafe’s own. Jev is TypeSafe’s first and so far only model under it.

Specialized decision models

Models trained for decisions, built on an existing open language model.

  • Microsoft

    Microsoft-Decision-1

    AvailableProduct modelannounced 9 Oct 2026
    Purpose
    Scores given answer options: yes/no, multiple choice, rating, rubric-based grading.
    What is notable
    According to Microsoft post-trained from the open model Qwen3.5-9B; returns a probability for each option and no rationale.
    Why it is here
    A large platform provider offers a decision model of its own as a product.
    Access
    In Microsoft Foundry, listed there as generally available.
    • Vendor claimMicrosoft tested the model in a comparison of its own across 36 benchmarks, which the post does not name.
    • LimitationAccording to the model card not designed as the sole automated decision-maker in consequential decisions about people.

    Sources: Microsoft Command Line: Microsoft-Decision-1 (9 October 2026), Microsoft Foundry model catalog: Microsoft-Decision-1 · checked 10 Oct 2026

    Model page in the registry: Microsoft-Decision-1 →
  • H2O.ai

    H2O-Lightning-4B

    AvailableSpecialized variantdate not stated by the source
    Purpose
    Choice, yes/no and score over text and images.
    What is notable
    According to the model card fine-tuned from Qwen3.5-4B in three stages; one decision is one forward pass with one output token.
    Why it is here
    A small open decision model that also reads images.
    Access
    Open weights under Apache-2.0, for self-hosting.
    • LimitationAccording to the model card evaluated mostly in English; decides best with up to about 16 options.

    Sources: Hugging Face: h2oai/h2o-lightning-4b model card · checked 10 Oct 2026

    Model page in the registry: H2O-Lightning-4B →
  • Chinh Nguyen

    Quyet-1.0-Large

    AvailableSpecialized variantdate not stated by the source
    Purpose
    Choice, score and yes/no with probabilities.
    What is notable
    According to the model card Gemma-4-31B-it with a merged LoRA fine-tune; the answer is read out as the letter of the chosen option.
    Why it is here
    Shows how a large open model becomes a decision model with a narrow fine-tune.
    Access
    Open weights under Apache-2.0; published by an individual.
    • LimitationAccording to the model card English and Vietnamese; other languages with lower accuracy.

    Sources: Hugging Face: chinhnc/Quyet-1.0-Large model card · checked 10 Oct 2026

    Model page in the registry: Quyet-1.0-Large →
  • Invergent

    Surogate Rune 26B-A4B

    AvailableSpecialized variantdate not stated by the source
    Purpose
    Decisions over text and images with a very large context.
    What is notable
    According to the provider 26B parameters with 4B active per token; a context of 262k tokens. The model card states calibration figures of its own.
    Why it is here
    The provider itself publishes that the model is overconfident at its default setting, and how to correct it.
    Access
    Open weights under Apache 2.0.
    • Vendor claimThe provider's own measurement: at the default setting 73.9% accuracy at 86.4% mean confidence; at a higher decision temperature the calibration error falls from 12.5% to 2.2%.

    Sources: Hugging Face: surogate/rune-26b-a4b model card, Invergent: Rune · checked 10 Oct 2026

    Model page in the registry: Surogate Rune 26B-A4B →

Generalists behind a decision interface

A general-purpose model reached through an endpoint that returns typed decisions.

  • OpenAI

    OpenAI Decisions API

    AvailablePreviewModel + harnessdate not stated by the source · in comparisons also "GPT-6 Luna Decisions"
    Purpose
    Typed answers about text and images: check a condition, pick an option, assign a level.
    What is notable
    No model name of its own: a decision interface in front of a generalist. Whether separate weights sit behind it is not documented.
    Why it is here
    Shows the second route to decisions: a generalist behind a decision interface instead of a model of its own.
    Access
    Public beta in the OpenAI API; the only model behind it at present is gpt-6-luna.
    • LimitationOpenAI advises setting thresholds from labelled examples of one's own application.

    Sources: OpenAI API docs: Decisions guide · checked 10 Oct 2026

Research and comparison systems

Experiments and procedures that appear in comparisons; not products.

  • Strands Labs

    Strands-Decider 2B

    AvailableResearch modeldate not stated by the source
    Purpose
    Fast decisions inside agent workflows.
    What is notable
    The part of the base model that generates text is removed and replaced by a head that only scores the offered options.
    Why it is here
    Here the decision is not trained in but built in: the model can no longer generate text.
    Access
    Open source code and weights under Apache-2.0; published as an experimental project.
    • Vendor claimIts own statement: answers at a confidence of 0.9 or more are right about 95% of the time.

    Sources: GitHub: strands-labs/strands-decider, Strands Agents: Introducing Strands Decider · checked 10 Oct 2026

    Model page in the registry: Strands-Decider 2B →
  • krishna-gogineni-765

    deck-31B

    AvailableModel + harnessdate not stated by the source
    Purpose
    Reads decisions out of an unmodified language model.
    What is notable
    Nothing was trained: a frozen Gemma-4-31B-it with a read-out procedure for option letters.
    Why it is here
    As a control: what an unmodified model achieves with the right read-out, without any training.
    Access
    Open server code from a single GitHub account; it has no weights of its own.

    Sources: GitHub: krishna-gogineni-765/deck31b · checked 10 Oct 2026

No rankingThe order is the order of this list. agi.jetzt does not grade these models and does not compare them with one another.

What decision models are described by

These models need other descriptive fields than a general-purpose model. They matter because the answer is consumed by a machine: there, response time and probabilities that can be trusted are requirements of their own.

Output type
A choice, a yes/no decision, a score, a ranking, a probability distribution or a structured object.
Calibration
Whether the stated probabilities match how often the model is right. See below.
Latency
How long one decision takes. In a chain of twenty decisions every tenth of a second adds up.
Size
Parameters and, for mixture-of-experts models, the parameters active per token — where published.
Interface
Whether the answer is read out deterministically or drawn at random, and what the confidence value means.
Limits of the schema
How many options, how many levels, how much context a request may carry.
Benchmark coverage
Which comparisons include the model — and whether a result belongs to the model alone or to a model plus harness.

These figures stand side by side. agi.jetzt does not combine them into one score. Higher accuracy does not mean more AGI, and lower latency does not mean more intelligence.

The eight entries side by side

Descriptive fields from the providers’ own pages. In the order of the list above.

ModelKindReturnsBuilt onSizeInputStatus
JevTypeSafe AIProduct modelchoice, score, probabilitiesNot publishedNot publishedText onlyAvailable
Microsoft-Decision-1MicrosoftProduct modelyes/no, choice, score, probabilitiesQwen3.5-9B, post-trainedNot stated (base model: 9B parameters)Text onlyAvailable
H2O-Lightning-4BH2O.aiSpecialized variantyes/no, choice, score, probabilitiesQwen3.5-4B, fine-tuned4B parametersText and imagesAvailable
Quyet-1.0-LargeChinh NguyenSpecialized variantyes/no, choice, score, probabilitiesGemma-4-31B-it, LoRA fine-tune31.3B parametersTextAvailable
Surogate Rune 26B-A4BInvergentSpecialized variantyes/no, choice, score, probabilitiesGemma-4-26B-A4B-it, fine-tuned26B parameters, 4B active per tokenText and imagesAvailable
OpenAI Decisions APIOpenAIModel + harnessyes/no, choice, score, probabilitiesgpt-6-luna behind a dedicated endpointNot publishedText and imagesAvailablePreview
Strands-Decider 2BStrands LabsResearch modelyes/no, choice, score, probabilitiesQwen3.5-2B without its text head, with a pointer head1.9B parametersTextAvailable
deck-31Bkrishna-gogineni-765Model + harnessyes/no, choice, score, probabilitiesGemma-4-31B-it, unmodifiedAs the base modelTextAvailable

Calibration, in plain words

If a system says "80 per cent sure" across many comparable decisions, roughly 80 per cent of those decisions should actually be right. Then it is well calibrated. A model that says 95 per cent and is right in 70 per cent of cases is overconfident — and software that trusts its number will act where it should have asked.

Calibration is a property of many predictions together. It does not say that a single answer is correct.

Why it matters

  • RoutingSend a case to the cheap path only when the model is sure enough.
  • Autonomous agentsLet an agent continue on its own above a threshold and stop below it.
  • Risk thresholdsDemand more certainty for a refund of 5,000 euros than for one of 5.
  • Escalation to peopleHand the uncertain cases to a person instead of all of them or none.
  • Decisions in softwareA probability can be compared, logged and tested. A sentence cannot.

What has been published

  • TypeSafe states that Jev is trained for calibrated decisions, but publishes no calibration metric.

    Sources: TypeSafe docs: System One

  • Invergent reports for Rune, on its own measurements: at the default setting 73.9% accuracy at 86.4% mean confidence (calibration error 12.5%); read at a higher decision temperature, 2.2%.

    Sources: Rune model card

  • Strands Labs reports for Strands-Decider that answers at a confidence of 0.9 or more are right about 95% of the time.

    Sources: strands-labs/strands-decider

  • Microsoft states the aim for Microsoft-Decision-1 — a 90% prediction should be right about nine times out of ten — without a figure in the text of its post.

    Sources: Microsoft Command Line

  • An independent preprint on text annotation finds the tested decision model’s confidence better calibrated than the verbalized confidence of 16 of 19 language models, while three of them show a lower calibration error — and on one task high confidence at near-chance performance.

    Sources: arXiv:2609.24574

  • An independent audit of Jev measured, on questions that cannot be answered, a calibration error of 0.023 with an option to abstain and of 0.793 once that option was removed.

    Sources: jujumilk3/jev-calibration-audit

agi.jetzt calls no model well calibrated. The figures above come from different tasks and different measurements and cannot be compared with one another.

What the comparisons show — and what they do not

There is no agreed benchmark for this class yet. Two comparisons are available, and both have limits.

Microsoft compared its own model with others on 36 benchmarks that its post does not name. It is a provider’s comparison of its own product. H2O.ai notes on its model card that the accuracy figures are Microsoft’s own, on benchmarks Microsoft chose.

JevBench is a leaderboard run by Benchmark Heaven, independently of TypeSafe, as a hobby project. Self-hosted systems answer 1,500 decisions, 1,200 of them sealed; a complete rerun of the sealed part requires the operator’s own evaluation process.

TypeSafe’s own evaluations measure Jev inside workflows the company built, against reference answers produced by two large language models rather than by people. TypeSafe itself writes that some bias could exist.

For these reasons this page shows no table of benchmark values. GPT-6 Sol, which appears in Microsoft’s comparison, is a generalist used there as a baseline for latency — not a decision model.

Sources: Microsoft Command Line: Microsoft-Decision-1, H2O-Lightning-4B model card, Benchmark Heaven: JevBench methodology, TypeSafe: Workflow evals, TypeSafe: Introducing System One Models & Jev

Why is this relevant to AGI?

Systems that act on their own have to decide all the time: which tool, which route, good enough or hand over. A model that answers such questions quickly and says how sure it is can be a building block for that — for routing, for gating a risky step, for scoring a result inside an agent system.

What makes it interesting for AGI research is the combination of structured decisions, stated uncertainty and automation between machines. A decision model is not closer to AGI than a generalist: it solves a narrower task.

What this does not prove about AGI

A fast decision with a stated probability is a narrow skill. It shows nothing about planning, learning or open reasoning, and a decision model is not closer to AGI than a generalist.

Relevance to AGI does not mean demonstrated progress toward AGI. These models are not part of the frontier models, and nothing on this page enters AGI Progress.

Sources

All sources read on 10 Oct 2026.