Fastino’s GLiNER2.5 redefines information extraction with boundary prediction
Fastino just shipped GLiNER2.5, an upgrade that swaps out the old habit of enumerating every possible span in favor of predicting where entities begin and end. The shift removes the need for a fixed maximum entity width, unlocks 4,096-word documents, and keeps compute linear in sequence length for a fixed schema—all while adding joint entity-relation decoding and per-span attributes. Three Apache-2.0 checkpoints land on Hugging Face at 74 M, 194 M, and 287 M parameters, with local CPU, CUDA, or MPS inference via a single pip install.
A simpler math for extraction
Previous GLiNER models scored every candidate span by pairing each start position with every allowed width. GLiNER2.5 instead scores token boundaries: start, end, and inside probabilities, then uses a sparse proposal stage to pair the most promising starts and ends per query. No distance restrictions, no grid of widths—just two passes that stay linear in sequence length for a fixed schema and candidate budget. The team reports a 56.17 macro F1 across 16 zero-shot benchmarks, a modest jump over GLiNER2’s 56.09, but with a notable 24.75-point gain on XNLI.
What the new numbers unlock
Memory savings from dropping explicit span representations let GLiNER2.5 train on sequences up to 4,096 tokens, and the library ships helpers like extract_entities_long that map long spans back to character offsets in the original document. Span length is now effectively unlimited: a 40-word indemnification clause costs the same compute as a two-word name. Users can also declare entity types, typed relations, and structural rules, then decode entities and relations jointly from the same candidate pool.
Plug and play for any stack
Fastino offers three checkpoints—small (74 M), base (194 M), and large (287 M)—all under Apache 2.0 on Hugging Face. Local inference runs on CPU, CUDA, or MPS; no hosted endpoints are available yet, so deployment means self-hosting. A two-person team can run the 74 M model on a standard CPU box, while larger organizations get a privacy-preserving alternative to per-token LLM extraction for tasks like PII redaction, contract clause extraction, and clinical entity tagging with negation and dosage attributes.
Why it matters
GLiNER2.5 shrinks the gap between cheap small encoders and expensive LLMs for information extraction without sacrificing flexibility. By removing span enumeration, it cuts memory overhead, widens context windows, and streamlines joint decoding—making high-quality extraction accessible to startups and scalable for enterprises. The open weights and CPU-friendly checkpoints lower the barrier to deployment, signaling that boundary-first architectures could soon become the default for production extraction pipelines.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.