GliNER2 represents a significant advancement in structured information extraction from text, offering a lightweight yet powerful alternative to large language models for classic NLP tasks. While SpaCy once dominated the NLP landscape with its accessibility and robustness, the rise of LLMs like ChatGPT has shifted focus toward more general-purpose models. However, these often bring unnecessary computational overhead for specific tasks. GliNER2 fills the gap by combining entity recognition, relation extraction, and structured JSON output into a single, efficient framework that runs on CPU. The core innovation in GliNER2 is its schema-driven approach. Users define extraction tasks declaratively—specifying entity types, relation types, and structured fields—then execute them in one inference call. This unified interface simplifies workflows and improves consistency. The model fastino/gliner2-large-v1, used in this evaluation, demonstrates strong performance across multiple extraction types while maintaining efficiency. Entity extraction in GliNER2 benefits from customizable descriptions for each entity type. For example, defining “Person” as names including nobility titles helps the model distinguish relevant mentions from noise. In testing with a Wikipedia excerpt about Ada Lovelace, the model accurately identified key entities such as people, inventions (like the analytical engine), locations, and historical events. Relation extraction allows users to define custom relation types such as parent_of, married_to, worked_on, and invented. The model correctly captures relationships like Ada Lovelace being married to William King or working on Babbage’s invention. However, it shows sensitivity to label naming—despite identical descriptions, the alias relation was consistently detected while same_as was not. This suggests that label consistency and naming conventions matter significantly in relation extraction. One of GliNER2’s standout features is structured JSON extraction via the extract_json method. This enables users to define arbitrary fields with types and descriptions, pulling clean, structured data directly from text. In the test, a schema was created to extract person details including name, birth and death dates, aliases, and relationships. While the output was mostly accurate, limitations emerged: gender was not inferred (despite Ada being referred to as a daughter), and some relationships were incorrectly assigned—such as Charles Babbage and William King being marked as parent_of Ada. These errors highlight the model’s strength in direct extraction versus reasoning or inference. Despite these issues, the structured output is highly valuable for knowledge graph construction. By combining entity, relation, and structured data extraction in a single pass, GliNER2 streamlines the process of transforming unstructured text into a graph-ready format. The results can be imported into Neo4j using a dynamic Cypher query that creates nodes, relationships, and properties while preserving the original text as a source reference. The model’s ability to run on CPU makes it ideal for production environments where cost, latency, and scalability are critical. It’s particularly well-suited for use cases requiring schema compliance, such as knowledge graph ingestion, data pipelines, and enterprise information systems. In summary, GliNER2 marks a meaningful return to focused, efficient NLP tools. While LLMs excel at open-ended tasks, they are often overkill for structured extraction. GliNER2 offers a smarter, faster, and more resource-friendly alternative—especially when precision, consistency, and performance matter. Its evolution from GliNER1 reflects growing maturity in specialized NLP models, and it stands as a compelling choice for developers building knowledge graphs and data pipelines.