GLiNER2

GLiNER2 is a 205M parameter encoder model that unifies named entity recognition, text classification, and relation extraction into a single CPU-efficient forward pass.

GLiNER2 transitions beyond simple entity recognition by introducing a declarative, schema-driven interface for complex information extraction. Built on a bidirectional transformer architecture, this model handles four distinct tasks (NER, classification, relation extraction, and structured JSON parsing) simultaneously without the latency or cost of frontier LLMs. It maintains competitive zero-shot performance (matching GPT-4o in specific F1 benchmarks) while remaining small enough for local deployment on standard hardware. By consolidating fragmented NLP pipelines into one pip-installable library, GLiNER2 provides a fast, private alternative for transforming unstructured text into clean, hierarchical data.

https://github.com/fastino-ai/GLiNER2