GLiNER

GLiNER performs named entity recognition without fixed entity types: you supply arbitrary type labels at inference time and a compact bidirectional transformer encoder finds matching spans, a zero-shot approach that avoids retraining for every new schema. Beyond standard NER the framework covers streaming extraction, joint entity and relation extraction, PII detection, and multi-task token classification across more than 100 languages. Several architectures are available, including a bi-encoder that encodes text and labels separately so large entity type libraries scale, plus uni-encoder and decoder-based variants. Models stay small enough for CPU deployment on consumer hardware, and the project supports torch.compile, FP16 and INT8 quantization, ONNX export, and production serving through Ray Serve with dynamic batching. Created by Urchade Zaratiana and collaborators and released under the Apache 2.0 license, it suits teams that need custom extraction without LLM-scale inference costs.