fastino/GLiNER2.5-Decide

This 340M parameter English classification model evaluates intent, routing, and sentiment labels during a single forward pass. It eliminates the need for prompt templates or generated tokens while accepting arbitrary label sets at call time. The publisher claims it serves operational decisions like customer service routing, banking requests, and review sentiment analysis. Specific benchmark results show a 60.2% exact-match accuracy on held-out examples across 17 domains. Developers install the package via pip and load the model using the AutoExtractor class for local execution. Tasks are performed by passing text and a dictionary of candidate labels to the classifytext method. Single-label tasks return one string, whereas multi-label tasks return every label above the configured threshold. A single call can score several heads simultaneously, such as classifying document type and determining priority in one operation. The release is explicitly a specialist tool and does not handle reasoning, explanations, or open-ended questions. The source notes that this specific model is designed for English text, recommending a different variant for multilingual inputs. Users should verify that their specific use case falls within the operational decision scope defined by the publisher. Potential outputs shown in the documentation are illustrative examples of return shapes rather than guaranteed predictions for every input.