Fastino Releases GLiNER2.5 for Efficient Information Extraction

Fastino released GLiNER2.5, an information extraction architecture designed to bridge the gap between the cost-effectiveness of small encoder models and the flexibility of large language models. This new release replaces the traditional span enumeration method with boundary prediction. Instead of scoring every potential entity span against a predefined width grid, GLiNER2.5 directly scores the start and end points of an entity. This fundamental change eliminates the constraint of a maximum entity width, enables processing of up to 4,096-word contexts, and maintains linear computation complexity relative to sequence length for a fixed schema. The architecture also facilitates joint entity-relation decoding, allows for cross-task label constraints, and supports per-span attributes. In zero-shot evaluations across 16 benchmarks, the multilingual checkpoint of GLiNER2.5 achieved an overall macro F1 score of 56.17, a slight improvement over its predecessor, GLiNER2, which scored 56.09. Notably, GLiNER2.5 demonstrated a significant 24.75-point gain on the XNLI benchmark. Fastino has made three checkpoints of GLiNER2.5 available on Hugging Face under the Apache 2.0 license. These checkpoints vary in size, with 74 million, 194 million, and 287 million parameters. The release also includes instructions for local deployment via pip install "gliner2[local]" for Python 3.10 and later, supporting inference on CPU, CUDA, and MPS. Currently, no third-party inference providers host these checkpoints, meaning self-hosting is the primary deployment strategy. The GLiNER2.5 model is positioned as a deployable solution for various industries. For smaller teams, the 74 million and 194 million parameter checkpoints can run on standard CPU hardware, eliminating the need for a GPU budget. Larger organizations can utilize GLiNER2.5 as a fine-tunable, privately hosted alternative to per-token large language model extraction. Key industries that can benefit include legal and contract operations, healthcare and clinical documentation, financial services, insurance claims processing, customer support, and the development of AI safety tooling. Specific applications range from PII detection and redaction to contract clause extraction, building knowledge graphs for AI agent memory, routing AI agents and models, classifying guardrails, and performing clinical entity extraction with attributes like negation and dosage. The previous GLiNER models operated by enumerating all possible entity spans, which involved pairing every start position with every possible width and scoring each against the defined schema. This approach tied computational cost to the width dimension of the entities.