Request access

GLINER2.5 - NOW AVAILABLE

Open-source small language model for efficient information extraction

GLiNER2.5 replaces span enumeration with a new boundary-prediction architecture, unlocking unlimited span length, long-context extraction, joint entity-relation extraction, constrained classification, and span attributes. Still small, still no GPU required.

Use GLiNER2.5

Read blog

GitHub repo

45M+

Hugging Face

model downloads

Hugging Face model downloads

5.2K

GitHub

stars

5.2K

GitHub stars

1.1B+

End

users

1.1B+

End users

FEATURES

Five new capabilities with GLiNER2.5

Extract entities of any size, handle full-length documents, constrain label combinations, classify individual spans, and connect entities into graphs, all in a single call. Available in base, multilingual, and small versions.

New

Span length

Unlimited span length

Extract entities of any size, from full postal addresses to clause-length legal references, with no width limit to configure or work around.

See a demo

New

Constraints

Constrained classification

Classify across several tasks with declared rules that keep label combinations consistent, preventing contradictory predictions.

See a demo

New

Richer context

Span attributes

Classify any extracted span for a custom attribute, such as sentiment, severity, or negation, decoded in the same forward pass.

Classify any extracted span for a custom attribute, such as sentiment, severity, or negation, decoded in the same forward pass as the entity.

See a demo

New

Full documents

Long-context extraction

Process contracts, reports, and transcripts in a single pass, with native chunking that merges results back to the original offsets for anything longer.

See a demo

New

Knowledge graphs

Joint Information Extraction (IE)

Extract entities and relations as one connected graph, guaranteed to conform to your schema, for agent memory or knowledge bases.

See a demo

COMPARISON

Comparing GLiNER models

GLiNER2.5

Constraint-aware structured extraction

GLiNER2

Entity extraction & structured parsing

GLiNER

First-gen GLiNER - Launched in 2023

MODEL SIZE (PARAMETERS)

0.3B (multi), 0.2B (base), 74M (small)

0.3B (multi) or 0.2B (base)

90M (medium) or 50M (small)

GENERAL NER (FEW-NERD) - F1

52.37 (multi), 55.14 (base)

51.49 (multi), 47.22 (base)

—

NATURAL LANGUAGE INFERENCE (XNLI) - F1

62.30 (multi), 54.49 (base)

37.55 (multi), 49.01 (base)

—

NAMED ENTITY RECOGNITION

✓

TEXT CLASSIFICATION

✓ (plus cross-task constraints)

✓(single and multi-label)

—

STRUCTURED / JSON EXTRACTION

✓

—

RELATION EXTRACTION

✓ (more optimized)

✓

—

CROSS-TASK CONSTRAINTS

✓ (more optimized)

✓

—

LONG-DOCUMENT UTILITIES

✓

—

SPAN ATTRIBUTES

✓

—

INPUT CONTEXT

4.096 words

2.048 tokens

512 tokens

MAX SPAN LENGTH

No cap

8 words

12 words

OPEN-SOURCE LICENSE

Apache 2.0

FROM THE COMMUNITY

A growing community of developers building with GLiNER

Share your story

The GLiNER approach let us ship zero-shot hallucination typing in our LettuceDetect models. Users bring their own labels and a 300M encoder types error spans across prose, code, and tool output.
Ádám Kovács
CTO & Co-founder, KR Labs
GLiNER is one of the most practical models we’ve integrated into OpenMed. Its zero-shot approach enables flexible clinical extraction across CPUs, GPUs, Apple Silicon while keeping sensitive health text local.
Maziyar Panahi
Founder, OpenMed
I'm a big fan. I really love that it is open source and so I personally optimized it a lot to make it much much faster which was important for my use case.
Max Buckley
Former Head of Knowledge Research, Exa
I am building a chrome extension for PII redaction. The best part about GLiNER is the speed and the ease to use it the browser.
Hossein Kazemi
CEO, Xeco Labs

The GLiNER approach let us ship zero-shot hallucination typing in our LettuceDetect models. Users bring their own labels and a 300M encoder types error spans across prose, code, and tool output.
Ádám Kovács
CTO & Co-founder, KR Labs
GLiNER is one of the most practical models we’ve integrated into OpenMed. Its zero-shot approach enables flexible clinical extraction across CPUs, GPUs, Apple Silicon while keeping sensitive health text local.
Maziyar Panahi
Founder, OpenMed
I'm a big fan. I really love that it is open source and so I personally optimized it a lot to make it much much faster which was important for my use case.
Max Buckley
Former Head of Knowledge Research, Exa
I am building a chrome extension for PII redaction. The best part about GLiNER is the speed and the ease to use it the browser.
Hossein Kazemi
CEO, Xeco Labs

I’ve been working in the NER space for a long time, and I loved GLiNER from the first time I saw it. The architecture was simple, flexible, and immediately felt like something that could be pushed much further, so I decided to become one of its maintainers.
Ihor Stepanov
Co-founder, Knowledgator
GLiNER is great! I used it for redacting PII from text.
Pedro Probst
Machine Learning Engineer, Nubank
It is quite versatile. A quick run of GLiNER can be used as silver data labeling for a project, or it can be used to adapt to situation where we need NER on a corpus but the predefined list of the NER types is not known beforehand.
Hoan Nguyen
Head of Data Science, XOMAD
GLiNER hits a sweet spot larger LLMs can’t: production-grade zero-shot entity extraction at a fraction of the cost and latency. No prompt engineering - just a small model that does one thing exceptionally well. For structured extraction, the price-to-performance is hard to beat.
Henrik Albihn
AI Engineer, Strange Loop

I’ve been working in the NER space for a long time, and I loved GLiNER from the first time I saw it. The architecture was simple, flexible, and immediately felt like something that could be pushed much further, so I decided to become one of its maintainers.
Ihor Stepanov
Co-founder, Knowledgator
GLiNER is great! I used it for redacting PII from text.
Pedro Probst
Machine Learning Engineer, Nubank
It is quite versatile. A quick run of GLiNER can be used as silver data labeling for a project, or it can be used to adapt to situation where we need NER on a corpus but the predefined list of the NER types is not known beforehand.
Hoan Nguyen
Head of Data Science, XOMAD
GLiNER hits a sweet spot larger LLMs can’t: production-grade zero-shot entity extraction at a fraction of the cost and latency. No prompt engineering - just a small model that does one thing exceptionally well. For structured extraction, the price-to-performance is hard to beat.
Henrik Albihn
AI Engineer, Strange Loop

USE CASE

What teams build with GLiNER

Privacy

PII redaction and privacy compliance

GLiNER's most common use case: teams run it locally, in the browser, or on-prem to mask personal data before it leaves their infrastructure.

Domain-specific NER for RAG and search

Tag the entities that matter in your domain (product names, error codes, internal terms, etc.) so retrieval matches on meaning, not keyword overlap.

Healthcare

Biomedical and clinical entity extraction

Zero-shot NER where labeled data is scarce. OpenMed maintains 30+ GLiNER fine-tunes for disease, drug, and pathology extraction, used in clinical notes and literature mining.

Knowledge graphs

Knowledge graph construction

Extract entities and their relationships directly from text or documents. GLiNER2 outputs structured graph data to populate knowledge bases and agent memory.

Safety

LLM guardrails and safety classification

A fast, local safety layer. GLiGuard classifies prompts and responses for jailbreaks, toxicity, and PII leakage.

Multilingual

Non-English and low-resource NER

Extend GLiNER to a new language with a handful of labeled examples. Fine-tunes already cover Japanese, German, Wolof, Turkish, and more.

Documents

Structured extraction from documents

Turn unstructured text into typed JSON. Pull structured records from contracts, invoices, and financial reports without prompt engineering.

Data labeling

Data labeling and annotation

Train your own models faster. Label Studio integrates GLiNER for NER pre-annotation, so teams generate silver labels and humans review instead of labeling from scratch.

Building something with GLiNER? Share your story.

Share your story

Move beyond general purpose language models.

Specialized, open language models that are faster, cheaper, and more accessible for most teams.

Use GLiNER 2.5

Move beyond general purpose language models.

Specialized, open language models that are faster, cheaper, and more accessible for most teams.

Use GLiNER 2.5

Products

Fastino Labs

Connect

Fastino Inc. (“Fastino”) develops specialized AI models and provides APIs designed to support structured data extraction, classification, reasoning, and production AI workflows. Fastino is a technology company and does not provide legal, financial, compliance, or advisory services.

Any outputs, predictions, classifications, or decisions generated through Fastino models are based on the configuration, data, and implementation provided by the customer. Fastino does not control, verify, or guarantee the accuracy, completeness, or suitability of model outputs for any specific purpose. By using this website or Fastino’s models and services, you acknowledge that all content and outputs are provided for informational and operational purposes only and agree to our Terms of Use and Privacy Policy.

2026 Fastino Inc.