GLiNER2.5-multi-Decide

Fastino's 287M multilingual classification model in the GLiNER2.5 family. It takes text plus any label set passed at call time and scores those labels in a single forward pass, with no prompt template and no generated tokens.

It shares the typed-decision idea behind Jev: instead of generating text, it scores a supplied label set and returns the label or labels. The card documents a Python AutoExtractor interface for local inference, not a System One endpoint, so it is adjacent rather than a drop-in replacement.

Operational decisions: customer and banking intent, travel and clinic requests, review sentiment, document type, email and ticket routing, human handoff, agent completion, moderation, severity, urgency and spam, plus passage questions, book classification and ordinal scales.

Exact-match accuracy on fastino/fast-decisions, 17 domains with 300 held-out examples each and identical text and labels for every model: GLiNER2.5-Decide 60.2%, GLiNER2.5-Decide-1B 59.6%, JevK5 57.6%, this model 56.7%, SemIf 56.4%, GLiFormer 49.0% and Laya Router 46.6%.

The card says to use it when the text is not English and to use GLiNER2.5-Decide (340M) for English-only text. The benchmark suite itself is English, so the multilingual advantage is a design intent rather than a measured result on that table.

Run pip install "gliner2[local]" and then load the model with AutoExtractor.from_pretrained("fastino/GLiNER2.5-multi-Decide"). Calls use model.classify_text(text, {head: [labels]}) and return a label string or the multi-label set above the threshold.

Fastino, as part of the GLiNER2.5 classification family; the base model is fastino/gliner2.5-multi-v1 and the repository is Apache-2.0. The card links the GLiNER2 paper on arXiv (2507.18546) and the fastino-ai/GLiNER2 GitHub repository.