System One models return decisions, not text.

What is a System One model?

The name borrows the fast, intuitive System 1 versus deliberate System 2 distinction popularized by Daniel Kahneman in Thinking, Fast and Slow. In software, this category covers models built to make constrained judgments: select a route, rate urgency or estimate whether a condition holds. Your application supplies the context and allowed answers, then uses the result to decide what happens next.

TypeSafe introduced Jev as a System One model in September 2026. The category has expanded rapidly: Cloudflare released Clef on October 1, and Liquid AI announced d1 with vision on October 5. Jev is one model in this category; other providers have different architectures, request contracts and limits. See the TypeSafe introduction, Cloudflare release and Liquid AI release.

Three question types

Choice, Score and Noul are the TypeSafe names for three useful decision shapes. Several providers implement them directly; other models need a runtime or adapter. Finite-answer output alone does not establish API compatibility.

Examples are illustrative, not measured predictions. Contracts: TypeSafe question types.

What is noul in Jev? The yes/no question type explained · Using a decision model as a support-ticket classifier

Model comparison

Contract details worth checking

• Tev1-4B experimental: Selects an option letter with logprobs. A binary choice can express yes/no, but is not a native Noul probability contract. Logprobs are not calibrated confidence.
• Bosun 3.1 1.7B: The model card names this checkpoint 1.7B and reports 2,031,739,904 parameters for its pinned base. Typed answers require the Bosun runtime.
• Laya: Several checkpoints exist. The vendor distinguishes base zero-shot results from tuned results; choose and evaluate the actual checkpoint.
• Decisions API (Luna): OpenAI documents finite-answer decisions, but its announcement does not establish Jev question types or drop-in API compatibility.
• GLiNER2.5-Decide: The supplied inventory attributes this to Fastino; a surfaced model card is under RedHatAI. Exact release identity and attribution remain unverified.
• Xor 1.2: The model card identifies release/xor-1.2 separately from main. Pin the intended revision rather than assuming main is 1.2.
• djev: This source describes a structured-decision runtime over DiffusionGemma, not separately trained weights. The supplied inventory names a different author; attribution to that author is not verified.

Choose by scenario

What independent evaluation shows

The September 29, 2026 preprint Evaluating and Benchmarking the System One Model Jev by Tobias Deußer, Lorenz Sparrenberg and Rafet Sifa evaluates Jev on 37 datasets. It reports strong results on several classification and reasoning datasets, with weaker performance on low-resource languages, fine-grained or noisy labels, and rubric judgments. It also finds that a fixed 0.5 binary threshold can be poorly placed: tuning on training data improves micro-F1 on UNFAIR-ToS from 0.50 to 0.75.

Those findings concern a specific Jev version and evaluation, not every System One model. Test your chosen model on a held-out set, tune thresholds separately, include out-of-scope inputs and compare a trained classifier when you already have labeled data. Keep accuracy, calibration, latency and total cost as separate measurements.

Guides and update policy

Last reviewed: 2026-10-06. This edition lists 27 model, API and runtime entries; Clef and Clef-flash are separate rows. We retain reported candidates with explicit Verification pending labels, then verify each field against a vendor release, repository or model card. A pending row is a research lead, not a recommendation. This is not a claim of a complete census. New internal links need a published guide. Unverified rates and contracts remain explicitly marked.

Self-hosted System One models: what actually runs locally · Using a decision model as a support-ticket classifier · The JSON request and response format, explained · Jev API rates and VerdictKit playground plans · Try a decision model in the playground

Frequently asked questions