Full guide: Jev alternatives: hosted, open-source and local options compared
A model counts as zero-shot when the labels arrive with the request. You do not train it on “billing”, “returns” or “shipping” first. You pass those words in with the text and it picks. The term for the idea is zero-shot classifier. This entry is about the models and how to choose between them.
NLI encoders. The Hugging Face task page lists facebook/bart-large-mnli as a zero-shot classification model. Per its model card, the method treats your text as a premise and builds a hypothesis from each label, such as “This text is about politics”, then reads the entailment probabilities. It runs one pass per label, so cost grows with the label count.
Span models. GLiNER is an encoder that extracts entities of any type you name at request time. It is built for finding spans in text, not for choosing one label for a whole document.
Decision models. These take a question and a list of options and return a probability for each, in one pass. Jev is the first. Open-weights and vendor alternatives include Laya, Kev, Clef and GLiNER2.5-Decide. The Choice primitive is the classification question.
What to compare on:
- Accuracy on your own labels. Public benchmarks use someone else’s.
- Calibration. If you gate actions on probability, the numbers must be honest.
- Latency and price. Jev costs $0.042 per million input tokens. Laya, Kev and GLiNER2.5-Decide have open weights, so you pay for hardware.
- Label count limits. Jev allows 255 options per Choice.
The full list is in Jev alternatives. Measured results, including where trained classifiers win, are in Jev benchmarks. See also classification model.