Information extraction teams have long faced a trade-off: small encoder models are cost-effective but inflexible, while large language models offer flexibility at a steep per-document price. Fastino's release of GLiNER2.5 aims to close this gap by replacing traditional span enumeration with a boundary-prediction architecture. Instead of scoring every candidate span across a width grid, the model directly predicts where an entity starts and ends. This shift eliminates the need to set a maximum entity width, supports a 4,096-word context, and keeps computational cost linear in sequence length for a fixed schema. It also enables joint entity-relation decoding, cross-task label constraints, and per-span attribute extraction. On 16 zero-shot benchmarks, the multilingual checkpoint achieves an overall macro F1 of 56.17, slightly outperforming GLiNER2's 56.09, with a notable 24.75-point improvement on XNLI. Fastino has released three checkpoints on Hugging Face under the Apache 2.0 license, at 74M, 194M, and 287M parameters.
Deployment and Accessibility
Yes, Fastino offers three GLiNER2.5 checkpoints on Hugging Face under the Apache 2.0 license. Local inference is supported on CPU, CUDA, or MPS using pip install "gliner2[local]" (Python 3.10 or later). As of now, no cloud inference provider hosts these checkpoints, so self-hosting is the recommended deployment path. This makes it straightforward for teams to integrate GLiNER2.5 into existing pipelines, whether for on-premises processing or edge deployments, while keeping data privacy and operational costs in check.
via MarkTechPost