GLiNER

GLiNER is an open-source family of small language models for efficient information extraction.

​

GLiNER2.5

GLiNER2.5 replaces span enumeration with a boundary-prediction architecture. It supports spans of any length within the model context, long-context extraction, joint entity-relation extraction, constrained classification, and span attributes. It remains small and can run without a GPU.

• Unlimited span length — Extract full addresses, legal references, and other long entities without configuring a maximum span width.
• Constrained classification — Declare rules across classification tasks to prevent contradictory label combinations.
• Span attributes — Classify extracted spans by attributes such as sentiment, severity, or negation in the same forward pass.
• Long-context extraction — Process contracts, reports, and transcripts with a 4,096-word context and chunk longer documents while preserving offsets.
• Joint information extraction — Extract entities and relations as one schema-constrained graph for knowledge bases and agent memory.

​

Compare GLiNER generations

GLiNER2.5 is the recommended starting point for new extraction projects. Use GLiNER2 for existing schema-based integrations and the original GLiNER for open-vocabulary NER checkpoints built on the first-generation package.