Compare AI decision models

A decision model, also called a System One model, does not write free text. It receives a state and a typed question and returns a fixed result: a choice among given options, a yes or no, or a score with probabilities. Applications read that result directly, for example to route a ticket or pick the next action.

An LLM generates text token by token and can answer anything, which also means its output has to be parsed and checked. A decision model is limited to the options or schema you pass in, so its answer is always machine-readable. It does not explain, summarize or chat; for that you still need an LLM.

Standalone (26): Own weights; picks from fixed options or fills a fixed schema. Adapter (7): A LoRA adapter that turns an existing language model into a decision model. Runs only with its base model. Classifier (4): A small model that makes narrowly scoped operational decisions as classification.

35 of 37 models publish open weights, which you can run yourself. The license decides commercial use, so check it in the table. Jev and GLiDE publish no open weights and run through the provider's API.

Most are small: 11 have fewer than 1 billion parameters, 10 between 1 and 5 billion, and 9 more than 5 billion. 7 profiles publish no parameter count, mostly adapters, whose size depends on the base model.

11 models document image input, among them AutoTrust JEV-27B-VL, Clef and Clef-Flash. Use the "Understands images" filter in the table for the full list.

By the popularity index the leaders are AutoTrust JEV-27B-VL, AutoTrust JEV-27B and GLiNER2.5-Decide. Popularity measures attention and use, not decision quality; a shared benchmark across decision models does not exist yet.