Our summary
Hugging Face published this post on 10 November 2026. It explains what decision models are, starting from TypeSafe AI's Jev, which TypeSafe describes as a System One model, and then shows where they now fit in the Hugging Face ecosystem.
It covers four routes. Load a model with a Transformers decision pipeline. Serve a GGUF model through llama.cpp using the same /v1/systemone API as Jev. Find compatible models through a decision-model tag on the Hub. Or fine-tune a model, with a toy example that routes support requests.
The training example is deliberately tiny, five labelled messages, so it shows the moving parts rather than a recipe for a useful model. The post is a tooling walkthrough, not a benchmark, so it does not tell you which decision model to choose.
Key takeaways
- Open decision models can be served with Transformers or llama.cpp behind the same request shape that Jev's hosted API uses.
- Search the decision-model tag on the Hugging Face Hub to find compatible models, which do not all share one architecture.
- To fine-tune a choice question, each example pairs a request with its correct option and trains with ordinary cross-entropy over the options.
