Fastino's GLiNER2 extends GLiNER beyond NER

Fastino researchers Urchade Zaratiana and Mary Newhauser have described GLiNER2, the follow-up to GLiNER, an open model released in late 2023. GLiNER treated zero-shot named entity recognition (NER) as a fast matching task rather than a slow generative one. GLiNER2 keeps that approach and adds three more tasks: text classification, relation extraction and structured (JSON) data extraction.

The authors pitch the model as a component for AI agents. In their framing, entities and relations give agents context and help build knowledge graphs. Classification sends queries to the right workflow or sub-agent, and schema-mapped JSON output passes clean data to downstream APIs. They argue that handing these jobs to small, specialized models cuts latency and leaves larger models free for reasoning and planning.

Fastino reports two results. It says GLiNER2 has the highest average zero-shot text-classification accuracy among open-source baselines. It also says the model closely matches GPT-5 on overall NER performance on the CrossNER benchmark while running entirely on CPU. The company also says fine-tuned GLiNER2 models can match or beat fully supervised, domain-specific state-of-the-art systems many times their size. The excerpt does not give specific scores, model sizes or the baselines it compared against.

GLiNER2 uses a bidirectional encoder transformer. Like the original, it projects input text and target labels into a shared latent space. Unlike the original, it adds a declarative, schema-driven interface. The post's section heading says all four tasks run in a single forward pass. The available text stops before giving further architectural detail, licensing or release information.