- Published on
A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction
- Authors
- Name
- DeepHorus
Fastino, a technology company focused on efficient AI tools, has released GLiNER2.5, a new boundary-prediction architecture designed to overcome key limitations in information extraction. This release addresses the common trade-off between the low cost but rigidity of small encoder models and the flexibility but high expense of large language models. The new model fundamentally changes how entities are located within text, marking a significant update to the GLiNER framework.
The core innovation is the replacement of span enumeration with boundary prediction. Earlier models scored every possible candidate span against a fixed width, which limited the maximum length of extractable entities. GLiNER2.5 instead predicts start and end scores at token boundaries, allowing the model to identify entities of any length without a predefined cap. This approach keeps computation linear relative to sequence length.
This architectural shift enables several new capabilities. The model can now process contexts up to 4,096 words, a substantial increase for long-document analysis. It also facilitates joint entity and relation extraction, where the model builds a globally consistent knowledge graph in a single pass. Furthermore, it supports constrained classification, where rules can bind labels across tasks to ensure logical consistency in the output.
Fastino has made three GLiNER2.5 checkpoints available on Hugging Face, with 74M, 194M, and 287M parameters. They are released under the permissive Apache 2.0 license, allowing for broad commercial and research use. Performance benchmarks show the multilingual model achieves a 56.17 macro F1 score across 16 zero-shot tasks, demonstrating competitive accuracy.
The practical implications are notable for deployment. The smaller 74M and 194M parameter models can run efficiently on a standard CPU, making powerful information extraction accessible to teams without dedicated GPU resources. This lowers the barrier to entry for startups and smaller organizations looking to implement extraction pipelines.
Potential applications span numerous industries. Use cases include PII detection and redaction, contract clause extraction for legal teams, clinical entity extraction in healthcare, and building knowledge graphs for AI agent memory. The model provides a self-hostable, cost-effective alternative to per-token API calls from large language models.
Future outlook points toward wider adoption as self-hosting becomes more common. Fastino’s release of a fully featured library with local inference support for CPU, CUDA, and MPS suggests a focus on practical developer experience. By removing key technical constraints, GLiNER2.5 positions itself as a versatile tool for structured data extraction tasks.
