GLiNER is a Named Entity Recognition (NER) model capable of identifying any entity type in a zero-shot manner. This architecture combines:

• An encoder for representing entity spans
• A decoder for generating label names

This hybrid approach enables new use cases such as entity linking and expands GLiNER’s capabilities. By integrating large modern decoders—trained on vast datasets—GLiNER can leverage their richer knowledge capacity while maintaining competitive inference speed.

---

Key Features

• Open ontology: Works when the label set is unknown
• Multi-label entity recognition: Assign multiple labels to a single entity
• Entity linking: Handle large label sets via constrained generation
• Knowledge expansion: Gain from large decoder models
• Efficient: Minimal speed reduction on GPU compared to single-encoder GLiNER

---

Installation

Update to the latest version of GLiNER:

# until the new pip release, install from main to use the new architecture
pip install git+https://github.com/urchade/GLiNER.git

---

Usage

If you need an open ontology entity extraction use tag label in the list of labels, please check example below:

from gliner import GLiNER

model = GLiNER.from_pretrained("knowledgator/gliner-decoder-large-v1.0")

text = "Hugging Face is a company that advances and democratizes artificial intelligence through open source and science."

labels = ["label"]

model.predict_entities(text, labels, threshold=0.3, num_gen_sequences=1)

If you need to run a model on many text and/or set some labels constraints, please check example below:

from gliner import GLiNER

model = GLiNER.from_pretrained("knowledgator/gliner-decoder-large-v1.0")

text = (
    "Apple was founded as Apple Computer Company on April 1, 1976, "
    "by Steve Wozniak, Steve Jobs (1955–2011) and Ronald Wayne to "
    "develop and sell Wozniak's Apple I personal computer."
)

labels = ["person", "company", "date"]

model.run([text], labels, threshold=0.3, num_gen_sequences=1)

---

Example Output

[
  [
    {
      "start": 21,
      "end": 26,
      "text": "Apple",
      "label": "company",
      "score": 0.6795641779899597,
      "generated labels": ["Organization"]
    },
    {
      "start": 47,
      "end": 60,
      "text": "April 1, 1976",
      "label": "date",
      "score": 0.44296327233314514,
      "generated labels": ["Date"]
    },
    {
      "start": 65,
      "end": 78,
      "text": "Steve Wozniak",
      "label": "person",
      "score": 0.9934439659118652,
      "generated labels": ["Person"]
    },
    {
      "start": 80,
      "end": 90,
      "text": "Steve Jobs",
      "label": "person",
      "score": 0.9725918769836426,
      "generated labels": ["Person"]
    },
    {
      "start": 107,
      "end": 119,
      "text": "Ronald Wayne",
      "label": "person",
      "score": 0.9964536428451538,
      "generated labels": ["Person"]
    }
  ]
]

---

Restricting the Decoder

You can limit the decoder to generate labels only from a predefined set:

model.run(
    text, labels,
    threshold=0.3,
    num_gen_sequences=1,
    gen_constraints=[
        "organization", "organization type", "city",
        "technology", "date", "person"
    ]
)

---

Performance Tips

Two label trie implementations are available. For a faster, memory-efficient C++ version, install Cython:

pip install cython

This can significantly improve performance and reduce memory usage, especially with millions of labels.