Today we’re launching GLiNER2.5, the most significant upgrade to the GLiNER model architecture to date. Across a suite of 16 benchmarks spanning diverse classification and extraction tasks, GLiNER2.5 achieves higher overall average F1 scores in comparison to GLiNER2, including a significant 24.75-point gain on XNLI. The gains come from a new architecture that predicts entity boundaries directly rather than enumerating every possible span. That change allows inference to scale linearly with document length and removes the max-entity-width cap for longer span length. Still built for entity extraction and classification, GLiNER2.5 adds five new capabilities: • Long-Context Extraction - full documents in a single call, so no information gets missed • Unlimited Span Length - entities of any length, including addresses, clauses, and titles are captured whole instead of cut off • Joint Information Extraction - entities and relations are decoded together for construction of unified, coherent knowledge graphs • Constrained Classification - labels stay logically consistent across tasks to avoid contradictions • Span Attributes - extracted spans come back with additional context in the same pass GLiNER2.5 comes in three variants: • GLiNER2.5-small (73.9M parameters) • GLiNER2.5-base (0.2B parameters) • GLiNER2.5-multi (0.3B parameters) Every day this week, we're releasing an interactive demo for each of the five new capabilities. Model weights are live on @huggingface under Apache 2.0: Read the full release blog: