Most of this site is about models that write text. Decision models do something different: they read some text and return a judgment, such as a label, a probability or a ranking, instead of a paragraph.
What kinds there are
- Safety and guard models check whether a prompt or a reply is harmful, and say which policy it breaks.
- Classifiers sort text into categories: topic, intent, sentiment, language.
- Rerankers score how well each candidate passage answers a question. They sit behind search and document chat to put the best passages first.
- Judges and reward models grade another model’s answer, for evaluation or for picking the best of several.
- Span extractors (such as GLiNER) find names, dates and other pieces inside a text.
Why they are interesting
- Small and fast. They often run in a small share of the memory of a chat model, and answer in milliseconds, so they run on modest hardware, including without a graphics card (CPU-only).
- Software-friendly answers. A probability or a label can be used directly by code: block above 0.9, route to the billing team, show the top three passages. No parsing of prose.
- A good fit for local. Classification of private text, such as tickets or documents, is a natural job to keep on your own machine (privacy).
Reading the ratings here
Because they are measured in latency, not tokens per second, these models are rated on how quickly they return an answer. The grade and Runs labels work as everywhere else (ratings). Libraries and servers that run them, such as Sentence Transformers, GLiNER, Infinity and Text Embeddings Inference, are on the apps page.