Plenty of AI work inside software is not writing but choosing: which team should get this support ticket, how urgent is it, is the customer asking for a refund. A new class of decision models answers those questions directly with probabilities instead of generating text, and Hugging Face has now given them a common interface across its tools. For developers, that means a router or classifier built on one decision model should be swappable for another without rewriting the glue code.
The interface is simple. You send a state, the information to judge, as text or structured JSON, plus named questions. Each question uses one of three primitives: choice, which picks among options and returns a probability for each; score, which rates something on an ordered scale and returns the expected level; and noul, a yes-or-no question answered with the probability that the answer is yes. Choice and score answers also carry a confidence figure, scaled so that an even spread across options is 0 and full certainty is 1. In Hugging Face's billing example, the model put 0.998 on billing, which works out to a confidence of 0.997.
The plumbing is the news. Transformers now has a DecisionPipeline, AutoModelForDecision and DecisionConfig, and transformers serve exposes a /v1/systemone endpoint. llama.cpp, which we review, serves decision models through the same endpoint, so a request written for a model on Hugging Face also works against a local GGUF model. The Hub has a decision-model tag, and models can be fine-tuned with the ordinary Trainer, with a choice question trained as plain cross-entropy over the options. The post notes that the models behind the interface vary, from encoder models such as ModernBERT to readouts on large generative backbones.
The context matters. TypeSafe AI's Jev, which we review, popularised the idea under the name System One model, and Google has just launched its own Gemini Decisions API with the same three question types. A shared open format makes it easier to compare these against each other and against prompting a chat model for JSON, which, as the post points out, can break the schema or answer outside the allowed options.
The limits are practical. A standard interface does not make any model accurate or calibrated, and the confidence figure is a rescaled probability, not a guarantee. As one commenter notes, the hard part in production is choosing thresholds and checking them against real outcomes that arrive later. Whether other model makers adopt this format, rather than keeping their own APIs, will decide how much it matters.