Hugging Face introduces decision models, a new model category

Hugging Face announced a new model category called decision models on November 10, 2026, introducing a way for models to output probability scored decisions over a fixed set of allowed answers instead of generating free text. The announcement appeared in Hugging Face's blog post on decision models, which also detailed support for the new category across the company's open source tools.

The concrete changes

• The interface is defined as State plus Questions plus Allowed Answers, producing Decisions and Probabilities.
• Three primitives are defined: Choice (a probability distribution over multiple options), Score (an expected value over ordered scales) and Noul (a binary yes or no).
• The Transformers library now ships the DecisionPipeline, AutoModelForDecision and DecisionConfig classes.
• llama.cpp adds support for decision models through a new /v1/systemone API endpoint.
• A new decision-model tag was added to the Hugging Face Hub for model discovery.
• TRL is adding first class training support for decision models.

What it means for people building with AI

Hugging Face positions decision models as a standard, lower overhead way to get constrained, schema compliant decisions from a model for classification and routing tasks, without relying on constrained decoding from a full generative LLM. The company notes decision models can be built on varied backbones, including encoder models such as ModernBERT as well as larger generative model backbones, giving developers a choice of tradeoffs between size and capability for a given routing or classification task.

Sources

• Decision Models in Hugging Face (Hugging Face, November 10, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.