Fastino Releases GLiNER2.5: Boundary-Prediction NER Architecture That Eliminates Span Enumeration

Fastino has open-released GLiNER2.5, a new named entity recognition architecture that replaces the conventional span enumeration approach with direct boundary prediction, significantly reducing computational overhead for information extraction tasks. Traditional NER models enumerate all possible spans up to a maximum length and score each — an approach that scales quadratically with sequence length. GLiNER2.5's boundary-prediction design instead predicts entity start and end positions directly, enabling faster inference and better performance on long documents and dense entity distributions. For developers building document processing pipelines, contract analysis tools, or any NLP system requiring structured extraction, GLiNER2.5 offers a drop-in improvement over prior GLiNER versions with lower latency at longer context lengths. The release continues Fastino's pattern of publishing production-grade, open NER tooling that outperforms general-purpose LLM extraction on structured entity tasks.