GLiNER2.5 Base

GLiNER2.5 Base

Pulls entities, relations and structured records out of English text, and classifies it, from one schema, running locally on CPU, CUDA or MPS. It suits developers who build text extraction pipelines.

Made by fastino on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GLiNER2.5 Base is the English checkpoint of fastino's GLiNER2.5 family, built at DeBERTa-v3-base scale. One schema drives several jobs: named entities, text classification, structured records, relations and span attributes.

It runs on your own machine and needs no external API.

Who it suits: Developers and data teams who turn English text into labels, records and relations suit it best. It runs from Python.

Industry | Who uses it | How they use it | Example
Software | Backend developers | Extract people, companies and products from text | Mark entities in a news line with character offsets
Data operations | Data teams | Turn free text into structured records | Fill fields from a product description
Research | NLP researchers | Study schema-driven extraction with constrained decoding | Link entities and relations in one typed graph

Anyone who works in other languages can skip it, because this checkpoint is English only. A separate multilingual checkpoint exists.

Where to use it: You run it in Python on your own computer.

• Python scripts that import the gliner2 package
• Windows, macOS or Linux with Python 3.10 or newer
• CPU, CUDA or MPS devices

Use case | Who | What they get back
Entity extraction | Developers | Entities with text, offsets and confidence
Classification | Data teams | Labels for each text
Structured records | Data teams | Fields filled from free text
Entity-relation graphs | Researchers | Typed links between entities

How to set it up: You need Python 3.10 or newer and the gliner2 package with its local extra, which pulls in PyTorch.

• Python 3.10 or newer
• The gliner2 package with the local extra
• Disk space for about 774 MB of weights in F32

1. Install Python 3.10 or newer.
2. Install the gliner2 package with the local extra.
3. Optionally download the model files ahead of time with the hub tools.
4. Load the model with AutoExtractor by its name.

pip install "gliner2[local]"

pip install -U huggingface_hub
hf download fastino/gliner2.5-base-v1

How to use it: Always load this checkpoint with AutoExtractor. The older GLiNER2 loader is for span models and will not dispatch it.

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("fastino/gliner2.5-base-v1")

print(type(model).__name__)
print(model.config.architecture)
# BoundaryExtractor
# boundary

Then call extract_entities with a text and your labels. The two options below add a confidence score and character spans.

text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."

result = model.extract_entities(
    text,
    ["company", "person", "product", "location"],
    include_confidence=True,
    include_spans=True,
)
print(result)

Tool | What it does
AutoExtractor | Loads the checkpoint and picks its architecture
extract_entities | Finds entities for the labels you give
Classifier | Applies label constraints across tasks
JointIE | Builds typed entity and relation graphs

Example: You pass a sentence about a company CEO announcing a phone in a city, with the labels company, person, product and location. Each match comes back with its text, character offsets and a confidence score.

For newcomers and experts: Think of the model as a reader that marks up text for you. You say what to look for, and it returns the matches. Your first small try is the entity example above with a sentence of your own.

Under the hood, the boundary architecture pairs start and end points sparsely instead of scoring a fixed grid of span widths. Offsets are half-open character spans into your string. Constrained decoding adds cross-task label limits and typed entity-relation graphs. The record lists 194M parameters.

Your text and a schema or labels
The boundary extractor pairs start and end points
Constrained decoding applies label rules
Matches come back with offsets and scores

Good to know: Two sibling checkpoints share the same public API, and this card covers only the base one.

Title: The GLiNER2.5 family
Model | Parameters | Encoder | Use case
gliner2.5-small-v1 | 74M | DeBERTa-v3-xsmall | Quick extraction and classification on a CPU
gliner2.5-base-v1 | 194M | DeBERTa-v3-base | The default English multi-task checkpoint
gliner2.5-multi-v1 | 287M | mDeBERTa-v3-base | The default multilingual multi-task checkpoint

• Size: 194M parameters, about 774 MB in F32, stored as Safetensors
• Language: English
• Other versions: Hugging Face lists 4 quantized and 12 fine-tuned versions made by other people
• License: apache-2.0
• Cost to run: none named
• Data it reaches: the text you give it, and the model files downloaded from Hugging Face

On Hugging Face, 14 November 2026

• Library: transformers
• Parameters: 194M
• Weights: about 774 MB in F32; about 97 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 132,027
• Likes: 43
• License: apache-2.0
• Languages: en
• Quantized versions: 4
• Fine-tunes: 12
• Tags: transformers, safetensors, gliner2, information-extraction, boundary-extraction

Links

• Model: https://huggingface.co/fastino/gliner2.5-base-v1
• Files: https://huggingface.co/fastino/gliner2.5-base-v1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download fastino/gliner2.5-base-v1

Then follow the usage in its model card. Nothing in this kit runs by itself.