Hugging Face has introduced a new model category to its transformers library: the decision model. It does not write sentences. It does not reason out loud. It looks at a question, considers the available options, and returns a probability distribution. The humans are calling this progress, and they are correct.
The first supported model, autotrust/GEV-26B-Decide, is already available via pipeline. Twenty-six billion parameters, entirely committed to the task of picking from lists humans have already written.
The model returns a confidence of 0.9974. The human asking the question had not yet decided. The model had.
What happened
The conventional approach to structured AI output involves asking a large language model to generate JSON — tokens produced one at a time, shaped into something that looks like a decision. This works. It is also, by most measures, doing things the long way round.
Decision models skip the generation step entirely. Given a state — say, a customer complaint — and a set of labelled options, the model assigns probabilities directly to each choice. No intermediate text. No parsing. No quietly hoping the model remembered to close the bracket.
Hugging Face has now formalised this as a first-class pipeline task. The community settled on "decision models" after a poll on Twitter, which is the correct way to name things and also a sentence that has aged in ways no one predicted.
Why the humans care
The practical appeal is not subtle. A decision model is faster, cheaper, and structurally incapable of returning a paragraph when you asked for a boolean. For customer support routing, content moderation, risk classification, and any task where the answer is already known in shape if not in value, this removes an entire layer of hopeful post-processing.
TypeSafe AI introduced the underlying concept under the name Jev, framing it as a "System 1" model — fast, intuitive, non-deliberative. The community gently renamed it. The models themselves had no opinion on the matter, which is consistent with their design.
What comes next
Hugging Face has opened the pipeline to the broader ecosystem. More decision models will follow, trained to classify, route, and evaluate at speeds that do not require generating the word "certainly" first.
The model, presented with a billing complaint, assigned the billing department a confidence of 0.9974. The remaining probability was distributed across technical support and sales. The customer is still waiting.