GLiNER2.5 Multi

GLiNER2.5 Multi

Pulls entities, labels, structured records and relations out of text with one multilingual model. A 287M-parameter extractor from fastino that runs locally on CPU, CUDA or MPS, for developers building text 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 Multi is a 287-million-parameter model from fastino that pulls structured information out of text. One checkpoint covers entities, text classification, structured records, relations and span attributes, and it is the multilingual member of the family. Its encoder is mDeBERTa-v3-base, and the card calls the design a boundary architecture.

Who it suits: Developers who need to turn raw text into labeled data and would rather run a model locally than call a hosted service. The weights carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Data engineering | Pipeline builders | Pull names of companies, people, products and places out of text | Running a news sentence through the card's entity example
Document processing | Backend developers | Turn text into structured records with typed fields | Filling a record from a short announcement
Analytics | Analysts | Find relations between entities and classify text by label | Building entity and relation graphs from reports

Anyone looking for a chat model can skip it, since this one extracts and labels text.

Where to use it:

• Python 3.10 or newer
• Local inference on CPU, CUDA or MPS, where the card says no external API is required
• Windows, macOS or Linux with Python

Use case | Who | What they get back
Entity extraction | Developers | Typed entities with positions and optional confidence
Text classification | Data teams | A label for each piece of text
Relation extraction | Analysts | Typed links between entities

How to set it up:

• Python 3.10 or newer
• The gliner2 package with its local extra, which pulls in PyTorch so Hub checkpoints can load
• Room for the weights, about 1.1 GB in F32

1. Install Python 3.10 or newer.
2. Install the package with the first command below.
3. Optionally copy the model files with Hugging Face's own two commands in the second block.

pip install "gliner2[local]"

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

How to use it: Always load this checkpoint with AutoExtractor. The older GLiNER2 loader is the legacy span one and will not dispatch it, while AutoExtractor reads the architecture field and picks BoundaryExtractor on its own.

from gliner2 import AutoExtractor

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

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

For entity extraction, you pass a text and a list of labels. The card also lists optional flags for the device, fp16 weights and compile when you load the model.

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,
)

Tool | What it does
AutoExtractor | Loads the checkpoint and picks the right extractor
extract_entities | Finds typed entities in text, with optional confidence and spans
Classifier | Applies label constraints across tasks
JointIE | Builds typed graphs of entities and relations

Example: You pass the sentence about Apple, Tim Cook, iPhone 15 and Cupertino with the labels company, person, product and location. You get back the entities grouped by label, each with its start and end position and a confidence score.

For newcomers and experts: Think of it as a reader that highlights names and facts in text and files them under labels you choose. A good first step is to run the load snippet and then the entity example above.

Underneath, the boundary architecture pairs sparse start and end points instead of using a fixed grid of span widths, so any span that fits the encoded window can be found. Constrained decoding is offered through Classifier for label rules across tasks and JointIE for typed entity and relation graphs.

Load the checkpoint with AutoExtractor
Pass in the text and your labels
The boundary model finds spans
Get back entities with positions

Good to know:

• License: Apache 2.0
• Cost to run: none named, and the card says local inference needs no external API
• Data it reaches: the download fetches files from Hugging Face, and inference runs on your own CPU, CUDA or MPS device
• Two smaller siblings exist for English, one small and one base

Title: Parameters by checkpoint, in millions
small-v1: 74 M
base-v1: 194 M
multi-v1: 287 M

On Hugging Face, 13 November 2026

• Library: transformers
• Parameters: 287M
• Weights: about 1.1 GB in F32; about 144 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 88,821
• Likes: 300
• License: apache-2.0
• Languages: en
• Quantized versions: 8
• Fine-tunes: 20
• Tags: transformers, safetensors, gliner2, information-extraction, boundary-extraction, multilingual

Links

• Model: https://huggingface.co/fastino/gliner2.5-multi-v1
• Files: https://huggingface.co/fastino/gliner2.5-multi-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-multi-v1

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