Classify text in 30+ languages into labels you choose, with no training. Runs multilingual-zeroshot-small (141M, Apache-2.0) entirely in your browser. The first run downloads the model once, then it is cached.
Text (any language)
Labels (comma-separated, English works best)
Hypothesis template ({} is replaced by each label)
Single-label scores sum to 1 (like the transformers zero-shot-classification pipeline); multi-label
scores are independent. Label wording matters: descriptive labels ("asks for a refund") beat terse ones. See the
model card for benchmarks and limitations; the base
model (308M) is more accurate than this small one.