Fastino Launches GLiNER2.5 with Boundary-Prediction Architecture for Information Extraction

Fastino released GLiNER2.5 to bridge the gap between small, rigid encoder models and large, expensive language models. [1] The new architecture uses boundary prediction to score where an entity starts and ends, replacing previous span enumeration methods. [1] This update removes maximum entity width limitations and enables a 4,096-word context window. [1]

For a fixed schema, this approach keeps computational requirements linear relative to sequence length. [1] The boundary-prediction change also enables cross-task label constraints, per-span attributes, and joint entity-relation decoding. [1]

In testing, the multilingual checkpoint achieved an overall macro F1 score of 56.17 across 16 zero-shot benchmarks, compared to 56.09 for the previous GLiNER2 model. [1] The updated model also recorded a 24.75-point improvement on the XNLI benchmark. [1]

Fastino released three checkpoints—sized at 74M, 194M, and 287M parameters—on Hugging Face under the Apache 2.0 license. [1] Currently, the models require self-hosting for deployment, as no inference provider hosts the checkpoints. [1]

GLiNER2.5 offers a compelling, self-hosted alternative for enterprises that currently rely on expensive per-token LLMs for extraction tasks. By shifting to boundary prediction, Fastino allows teams to process significantly longer documents without computational costs growing exponentially. The incremental gain on overall zero-shot benchmarks is minor compared to GLiNER2, but the architectural unlocks and specific gains on XNLI represent a solid improvement for cross-lingual tasks. What the sources don't address: how the model's inference speed compares to other extraction-focused small models in real-world deployment scenarios.