Fastino Drops GLiNER2.5 With Boundary Prediction to Fix Context Limits

Fastino releases GLiNER2.5, replacing span enumeration with boundary prediction to process variable-length entities. The architecture shifts to predicting start and end boundaries, pushing context limits to 4,096 words while keeping computation linear. The multilingual checkpoint hits 56.17 macro F1 across 16 zero-shot benchmarks, scoring a massive 24.75-point gain on XNLI. Three models—74M, 194M, and 287M parameters—are live on Hugging Face under Apache 2.0, requiring self-hosted inference since no provider currently hosts them. Ditching the width grid for boundary proposals is a clever way to escape quadratic scaling, but raw infrastructure deployment is what you'll be paying for.