Fastino Launches GLiNER2.5 for Information Extraction

Information extraction workflows often force engineering teams into a difficult compromise. They can deploy compact encoder models that run cheaply but lack flexibility, or they can rely on large language models that adapt easily to new schemas while incurring high costs per document. Fastino has introduced version 2.5 of its GLiNER framework to bridge this gap. The software offers a middle ground, combining the efficiency of smaller architectures with the zero shot generalization typically associated with much larger generative systems.

The core technical shift in this release is the abandonment of span enumeration in favor of boundary prediction. Previous iterations of the model evaluated every possible text span against a permitted width grid, a process that tied computational overhead directly to entity length and imposed strict ceilings on how long an entity could be. The new architecture instead predicts start and end scores over token boundaries, paired with token level inside scores. A sparse proposal stage identifies promising start and end points for each query without distance restrictions, and a dedicated reranking head evaluates the resulting candidates.

This modification yields several performance advantages for developers working with unstructured text. Computation scales linearly with sequence length for a fixed schema, and the maximum context window expands to 4,096 words. Furthermore, the updated design supports joint entity and relation decoding, cross task label constraints, and per-span attributes. Relation candidates are drawn directly from the shared pool rather than requiring a separate processing pipeline, simplifying the overall extraction logic.

Fastino has published three model checkpoints on Hugging Face under the permissive Apache 2.0 license. These checkpoints are sized at 74 million, 194 million, and 287 million parameters. Unlike massive generative models that require cloud infrastructure, these files are small enough for local execution on standard central processing units, graphics processing units, or Apple silicon via standard Python installation commands. Developers must currently self-host the weights, as external inference providers do not yet host them.

Performance evaluations conducted across sixteen zero shot public datasets demonstrate steady improvements over the previous generation. The multilingual variant of GLiNER2.5 achieves an overall macro F1 score of 56.17, compared to 56.09 for its predecessor, while the base variant climbs to 54.87 from 53.34. The most notable performance leap occurs on the XNLI benchmark, where the multilingual model jumps by 24.75 points to reach 62.30. Additional gains appear on specialized benchmarks like Few NERD and across untrained languages such as Romanian RONEC.

Adopting the new release requires minor migration steps for existing users. Teams must load the architecture using the new AutoExtractor class rather than relying on legacy span loaders. Because no managed API endpoints currently exist, production deployments will require engineering resources for self-hosting and scaling local inference infrastructure across development teams.