GLiNER2 Base

GLiNER2 Base is a compact English-language encoder designed around the idea that one model should handle the bulk of practical text-extraction work. It is built on a DeBERTa-v3-base backbone wrapped in the GLiNER2 architecture, and its main novelty is consolidation: named entity recognition, zero-shot text classification, structured JSON extraction, and relation extraction all run through the same forward pass, so applications no longer need separate pipelines or chained models for each task. Because the model treats labels, relation types, and even output schemas as inputs at query time, developers can point it at a new domain by writing a new prompt rather than collecting labeled training data.

The practical profile of GLiNER2 Base is shaped by its small footprint. With roughly 205 million parameters, it is intended for fast CPU-based inference, privacy-friendly local processing, and lightweight serverless or edge deployments where avoiding GPU costs matters. The included use cases document processing, knowledge-graph construction, content tagging, and pulling structured fields out of unstructured documents such as invoices or resumes, which matches the model's schema-driven extraction strengths. Fine-tuning and deployment are surfaced through the associated Pioneer platform, and the open Python `gliner2` package lowers the barrier to integrating the model into production extraction pipelines.