Hugging Face has added a first-class "decision model" interface to its Transformers library, betting that a lot of enterprise AI does not need to generate text at all. Instead of prompting a language model to emit JSON, a decision model takes a state and a set of predefined options and returns a probability distribution over those options — no tokens generated.

The new pieces are DecisionPipeline, AutoModelForDecision and DecisionConfig, plus a decision-model tag on the Hub and a /v1/systemone endpoint that works both in Transformers and in llama.cpp. The interface is built around two inputs — a "state" and a set of "questions" — and three primitives: choice (pick one of several answers), score (place something on an ordered scale, returning the expected value) and noul (a Bernoulli yes/no probability). Choice and score also return a "confidence" that normalises how concentrated the model's belief is.

The company argues the approach fixes real problems with the "just ask the LLM for JSON" pattern: models may produce invalid JSON, return an out-of-range label, or waste compute generating a full response when the application only needed to choose between options. Decision models can be trained with the Trainer API or, for now, by hand; a TRL pull request for first-class training support is open. A quirk: with compatible autoregressive models you can try the interface without training anything, since the pipeline scores answer logits in a batched forward pass.

Hugging Face is explicitly not requiring a single architecture. The decision-model tag already spans encoder-based models such as ModernBERT and larger generative backbones, with named examples including TypeSafe AI's Jev line, GLiDE, H2O-Lightning, Kakao's v-decide and Drex DLM. The community settled on "decision model" over TypeSafe's earlier "System One" framing — a name about the model's computational role rather than a theory of cognition.