Fastino has released GLiNER2.5, an information extraction model that replaces span enumeration with boundary prediction. This architectural change allows for longer context windows up to 4,096 words and linear computation with respect to sequence length. The model supports joint entity-relation extraction, cross-task label constraints, and per-span attributes, offering a more flexible and efficient alternative to existing methods. Three checkpoints are available on Hugging Face, suitable for various deployment scenarios including CPU-based inference. AI
IMPACT Offers a more efficient and flexible approach to information extraction, potentially reducing costs for tasks like PII detection and knowledge graph creation.
RANK_REASON This is a release of a specific model for information extraction, not a frontier model release.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
