Fastino Releases GLiNER2.5: A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction

Fastino has released GLiNER2.5, a new model for information extraction that fundamentally changes how entities are identified in text. It shifts from scoring every possible text span to a more efficient boundary-prediction system, eliminating previous constraints on maximum entity length. This architectural update enables long-context processing and joint extraction of entities and relations.

Key takeaways

• Replaces span enumeration with a boundary-prediction architecture, removing hard limits on entity length.
• Enables joint entity and relation extraction with rules for building globally consistent graphs.
• Supports long-context extraction up to 4,096 words with helpers for processing documents.
• Released as three Apache 2.0 licensed models, including a multilingual version, designed to run on CPU.
• Target applications include PII detection, contract analysis, clinical data extraction, and building knowledge graphs.

The Core Architectural Shift: From Span Enumeration to Boundary Prediction

GLiNER2.5 replaces the previous method of enumerating and scoring every possible candidate span within a fixed width limit. Earlier models located entities by pairing every start position with every allowed width, a design that tied computation to a width axis and imposed a hard ceiling on entity length.

A New Scoring Mechanism

The new boundary-prediction architecture has the model predict start and end scores over token boundaries, plus inside scores over tokens, instead of scoring spans. This eliminates the hard ceiling on entity length present in earlier versions, allowing a span to open at the first token and close at the last.

Sparse Proposal and Reranking

A sparse proposal stage selects the most promising start and end positions per query, pairing them with no restriction on distance. A reranking head then scores each candidate using boundary evidence and span content. This change keeps computational cost linear in sequence length for a fixed schema and removes the computational tie to a maximum entity width.

Key Capabilities and Technical Specifications

The GLiNER2.5 architecture enables long-context extraction with a maximum sequence length of 4,096 words. The library includes native helpers for chunking long documents, such as extract_entities_long and JointIE.extract_long, which remap spans to character offsets in the original document.

Joint Extraction and Advanced Classification

The model provides joint entity and relation extraction, letting users declare entity types, typed relations, and structural rules like unique_head=True. A beam search then assembles a globally consistent graph, with invalid combinations never admitted. It also introduces constrained classification via C.implies and C.excludes rules that bind labels across tasks during decoding. Furthermore, per-span attributes, such as sentiment, can attach to specific entity types via applies_to and are decoded in the same forward pass.

Released Models

Fastino has released three checkpoints under the Apache 2.0 license on Hugging Face. The models are: gliner2.5-small-v1 (74M parameters, English), gliner2.5-base-v1 (194M parameters, English), and gliner2.5-multi-v1 (287M parameters, multilingual). All three share the same public API and are runnable on standard CPU boxes.

Performance Benchmarks and Deployment Practicalities

In a zero-shot evaluation across 16 public datasets, the multilingual GLiNER2.5 checkpoint achieves an overall macro F1 score of 56.17, compared to 56.09 for the previous GLiNER2 Multi. The most significant gain is on the XNLI benchmark, where performance jumps by 24.75 points.

You can deploy the model via local inference on CPU, CUDA, or MPS using pip install ‘gliner2[local]’ for Python 3.10+. No external inference provider currently hosts the checkpoints, making self-hosting the required deployment path. The smaller 74M and 194M parameter checkpoints are specifically made to run on standard CPU boxes, providing an accessible option for teams without a GPU budget.

Target applications for the technology include PII detection and redaction, contract clause extraction, clinical entity extraction with attributes like negation and dosage, building knowledge graphs for agent memory, and implementing AI safety guardrails.