FrontierAugust 24, 2026via MarkTechPost
Fastino Releases GLiNER2.5: A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction
Why it matters
A foundational shift in how NER models work: boundary prediction replaces the compute-heavy enumeration step that made extraction cost balloon with entity density. Practitioners building IE pipelines get faster, denser inference and open weights to build on.
Key signals
- Fastino GLiNER2.5 released with boundary-prediction architecture (replaces span enumeration)
- Three Apache 2.0 checkpoints: 74M, 194M, 287M parameters
- All checkpoints CPU-runnable
- Joint entity-relation decoding added
- Constrained classification support
- Span attributes capability
- 4,096-word context window
- Macro F1: 56.17 on 16 zero-shot benchmarks
- Entity width no longer scales compute cost with model size
- Fastino released GLiNER2.5 with boundary-prediction architecture (replaces span enumeration)
- Features: joint entity-relation decoding, constrained classification, span attributes, 4,096-word context
- Macro F1 56.17 on 16 zero-shot benchmarks
- Eliminates quadratic compute cost of candidate generation
The hook
GLiNER2.5 ditches span enumeration entirely—entity extraction now scales with text length, not entity count. Three Apache 2.0 weights, all CPU-runnable.
Fastino released GLiNER2.5, replacing span enumeration with boundary prediction so entity width no longer costs compute. Three Apache 2.0 checkpoints ship at 74M, 194M, and 287M parameters, all CPU-runnable. The release adds joint entity-relation decoding, constrained classification, span attributes…