Fastino Debuts GLiNER2.5, a Boundary-Prediction Extraction Model
Fastino has released GLiNER2.5, a new architecture for information extraction that replaces span enumeration with direct boundary prediction, according to a report from Marktechpost.
Traditional information extraction systems built on the GLiNER family typically work by generating candidate spans of text and scoring each one against target entity types, a process that scales poorly as document length and label counts grow. Marktechpost reports that GLiNER2.5 instead predicts the start and end boundaries of relevant spans directly, avoiding the combinatorial overhead of enumerating every possible span in a document.
This boundary-prediction approach is positioned as a more efficient alternative for tasks such as named entity recognition and structured data extraction, where systems must identify and label spans of text corresponding to categories like names, dates, organizations or custom user-defined labels. By removing the enumeration step, the architecture aims to reduce computational overhead while maintaining extraction accuracy, according to the report.
Marktechpost’s coverage frames the release as part of ongoing efforts in the open information extraction space to build models that generalize across arbitrary label sets without requiring task-specific fine-tuning, a hallmark of the broader GLiNER lineage of zero-shot and few-shot extraction models. Fastino’s release extends this line of work with an architectural change intended to improve scalability for production use cases involving long documents or large numbers of entity types.
Details on benchmark comparisons, model size, licensing terms and availability of the model weights were not fully specified in the available report. The release adds to a growing body of open-source tooling aimed at making structured information extraction more efficient for developers building applications on top of large language models and lightweight extraction pipelines.
Based on reporting by www.marktechpost.com.
