Information extraction teams must choose between cheap rigid encoders and costly flexible LLMs. Fastino’s GLiNER2.5 replaces span enumeration with a boundary‑prediction scheme: the model scores start and end token positions and token‑inside scores, then proposes sparse start‑end pairs and reranks them. This removes the maximum entity width, yields linear‑in‑length computation for a fixed schema, and enables a 4,096‑token context.
Three checkpoints (74 M, 194 M, 287 M parameters) are released on Hugging Face under Apache 2.0; they share a unified AutoExtractor API and support CPU, CUDA, or MPS inference via pip install "gliner2[local]" (Python 3.10+). Zero‑shot evaluation on 16 datasets shows the multilingual variant at 56.17 macro F1 (vs 56.09 for GLiNER2) and a 24.75‑point jump on XNLI (62.30 vs 37.55). Smaller variants also improve on Few‑NERD and Romanian RONEC.
- Context window: 4,096 tokens
- Parameter sizes: 74 M / 194 M / 287 M
- License: Apache 2.0
- Deployment: self‑hosted via local pip package (CPU, CUDA, MPS)
Why this matters
The shift from span enumeration to boundary prediction cuts the quadratic cost tied to entity width, allowing models to scale to longer documents without increasing compute per token. This architectural tweak also unifies entity and relation extraction, simplifying pipelines for applications that need joint schema‑aware predictions.
Share this article
Found this insightful? Share it with your community on Reddit, X, or copy the link.