Category explainer · independent
What is a decision model?
A decision model answers a question whose legal answers you declare before the call. It scores that fixed set in one pass and returns a typed value with a probability, instead of generating text token by token. Cloudflare and AWS both use "System One models" for the category, after the name TypeSafe gave the family that started it.
The category appeared in September 2026 and grew quickly. This page is the map: the defining property, the models that exist, the jobs they suit, and the limits that do not go away no matter which one you pick.
The defining property
With a generative model, the output space is open and the application inherits a parsing problem. With a decision model, the answer space is closed before inference: a list of options, a numeric scale, or two values. The model scores the candidates and returns a distribution over them.
That single design choice explains most of the category's advertised benefits. Nothing is emitted token by token, so there is no generation step to wait for. Nothing can drift out of format, so there is no schema repair path. And because the model is trained against proper scoring rules in several implementations, the probability is intended to be thresholded in code rather than read as a vibe.
The 2026 family map
Release dates cluster tightly: Jev on 15 September, Laya on 18 September, Clef and Strands Decider 2B both on 1 October 2026. Most of the published benchmark tables come from the vendors themselves, often running competitors on their own harness.
| Model | Maker | Weights | Where it runs |
|---|---|---|---|
| Jev (System One) | TypeSafe AI | Closed | Hosted API, plus third-party gateways |
| Clef / Clef-flash | Cloudflare | Apache 2.0 | Workers AI or self-hosted |
| Laya | Convai Innovations | Apache 2.0 | Self-hosted, pip installable |
| Strands Decider 2B | AWS Strands Labs | Apache 2.0 | Local CPU or GPU |
| Kev 4B and other gateway entries | Various | Varies | Listed on decision-model gateways; check each model page |
Gateway catalogs change quickly and mix open and closed entries. Confirm the licence and request shape on the model page before assuming two entries are interchangeable.
Where the model belongs in an application
- →Routing: choose a queue, a tool, or which model should handle a request.
- →Gates: approve or deny an action, flag a prompt for human review, block a risky tool call.
- →Triage and classification: intent, severity, language, or category from a fixed label set.
- →Scoring and ranking: urgency, relevance, factuality, or quality on a scale you define.
- →Verification: check a generated answer against a criterion before it ships.
The use cases section of this directory writes each of these up as a concrete arrangement of the three primitives.
Where it does not belong
- →Anything a person reads: drafts, summaries, explanations, chat replies.
- →Open-ended questions where you cannot enumerate the candidates.
- →Work that needs the reason alongside the verdict in the same response.
- →Long complex reasoning — AWS explicitly notes the single-pass design is worse at solving complex problems than reasoning models.
Frequently asked questions
Is a decision model just a classifier with better packaging?
+
Partly, and the criticism is fair. Laya is an encoder with a decision layer on top, and AWS built Strands Decider 2B by removing a language-model head and adding a pointer head. The practical differences are that questions are defined at request time rather than at training time, and that a probability is returned for every option rather than one best label.
Why are decision models faster than LLMs?
+
Because they produce no tokens. A generative model's latency scales with the length of what it writes; a constrained evaluation scores a fixed answer space in one pass. Published figures vary widely by harness — Cloudflare measured Clef-flash at 38.8ms median and Jev at 524.1ms in its own run, while an independent race through OpenRouter found Jev faster on every task. None of these is a controlled comparison.
Do decision models replace LLMs?
+
No. They give up text generation entirely. The common architecture keeps both: a generative model handles anything a human reads, and a decision model handles the high-frequency judgements that code consumes.
Is "System One" the same as System 1 from cognitive science?
+
It is a product name TypeSafe chose for its model family, and other vendors have adopted it as a category label. Do not read it as a claim about human cognition or as a technical standard. There is no standards body behind the term.
Where should I start?
+
Try a typed question in the browser playground first, then read the API request shape. If you need to run inside your own network, start with the open-weight options rather than looking for a self-hosted build of Jev.
Keep reading
Decision models vs LLMs
Output shape, parsing risk, latency, cost, and calibration side by side.
How a decision API works
Request shape, question types, and confidence thresholds.
Composition patterns
Routers, guards, triage, and verification written up concretely.
Choice, Score and Noul
The three question shapes the whole category is built on.
Sources
- Cloudflare — Introducing Clef, including a definition of decision models ↗
- AWS ML blog — Strands Decider 2B and the agent hot path ↗
- TypeSafe AI — documentation ↗
- Laya on PyPI ↗
Reviewed 2026-11-13. Vendor- and project-reported figures are labelled as such; none were independently reproduced by this site.