Decision Models

Built with decision models

More than 700 public repositories mentioning Jev appeared in the three days after launch. Ten of them, each checked against its own repo.

Building something? Send it to if@decisionmodels.ai and it goes on the list. Inclusion is not endorsement, and nobody pays to be here.

What is a decision model?

Decision model is the term this site uses for a system whose primary output is a bounded, machine-consumed judgement: a value from a declared answer space, usually with a score over the options. The implementation varies. It might be a classifier head, a direct readout of a language model’s logits, constrained decoding into a schema, or a purpose-trained model such as Jev.

On the name TypeSafe calls Jev a System One Model, after the fast-versus-deliberative distinction popularised by Daniel Kahneman. Their own AI primer also describes the work as training “decision models with calibrated probabilities instead of optimizing for generated text.” This site uses decision model in that descriptive sense. It is not a settled category name, and the phrase already means something specific in decision analysis and in DMN.

The interface is not new: what differs between the two paths below is how the answer is produced, not whether software can get a typed one.

A worked example

A customer message arrives, and the system needs a category — nothing more:

message: "My card was charged twice."

schema:
  category: ["billing", "technical", "sales", "other"]

A decision model returns the answer in the shape the software asked for. Illustrative output, not a measured model result:

{"category": {"value": "billing", "probability": 0.96}}

The value is already bounded by the declared answer space. Whether the probability beside it is useful is a separate question: a normalised score is not a calibrated one, and calibration has to be measured against representative data.

This did not start with Jev

When Jev launched, one of the loudest replies came from a GLiNER2 co-author, pointing out that scoring declared options in a single forward pass has had open weights for years. He is right, and it is worth being precise about what is actually new.

1. Loading…

What Jev claims that this lineage does not is calibration as an explicit training objective, plus the latency and price bundle. Those are vendor-reported and not yet independently reproduced, which is the gap this site intends to close.

Ecosystem

One hosted model, four you can run yourself, one router. Click a row for detail.

Hosted models, local experiments, adapters and tooling are separated because they are not peers. A dot means nobody has established that property, not that the answer is no. Claims are sourced to maintainer documentation and marked vendor-reported where nobody has independently reproduced them. Inclusion is not endorsement, and this is not a complete inventory of classifiers or structured-output models.

The three paths do not agree

One open model, one task, three ways of asking. The speed difference is the easy result. The disagreement is the one worth knowing about.

Model

Qwen3 0.6B · 751.63M · Q4_K_M

Runtime

Ollama 0.33.2

Machine

MacBook Pro · Apple M3 Pro

Cores

11 · 5P + 6E

Memory

36 GB unified

OS

macOS 26.4.1

Inputs

25 × 3 repeats × 3 paths

Temperature

0

Three warm-up rounds before timing, 225 timed calls. Schema validity was 100% on all three paths, so on this task constrained decoding bought nothing over parsing. This measures mechanics, not accuracy and not calibration: there are no ground truth labels here, so “disagreement” means the paths differ, not that one is right. Latency is local and says nothing about any hosted API. If you reproduce this, publish your hardware too. Raw results (225 rows) · Summary · The script and repo

python3 output-paths.py --model qwen3:0.6b --repeats 3

The two paths, axis by axis

Generalisations, not laws: capabilities vary by model, and the two approaches are complementary rather than mutually exclusive.

Use cases

Show all ten

Resources

Follow the decision-model ecosystem.

New models, benchmarks, APIs and open-source projects.

Double opt-in: we send one email, and nothing happens until you click the link in it. We store your email address and the times you signed up and confirmed, in a Cloudflare D1 database. Nothing else. No advertising. Reply to any email to be removed, or write to if@decisionmodels.ai.