Editorial illustration for Fastino Drops GLiNER 2.5, Replacing Span Enumeration With Boundary Prediction
GLiNER 2.5 Replaces Span Enumeration With Boundaries
Named entity extraction has run on the same trick for years: enumerate every possible span in a document, then score each one against a maximum width. It works, but it caps how long an entity can be and scales badly as documents grow. Fastino's answer, released this week, is GLiNER2.5, and it throws out span enumeration entirely.
Instead of scoring candidate spans, the model predicts where an entity starts and where it ends. No width grid, no cap on entity length, and a jump to 4,096-word context while keeping computation linear for a fixed schema.
That architectural swap isn't cosmetic. It opens the door to joint entity-relation decoding, label constraints that carry across tasks, and per-span attributes, capabilities that span enumeration made awkward or impossible. Three checkpoints, at 74M, 194M, and 287M parameters, are live on Hugging Face under Apache 2.0, positioned for teams stuck between expensive per-token LLM extraction and small encoders that can't flex. What the boundary-prediction switch buys in raw accuracy, and where it falls short, shows up in Fastino's own benchmark numbers against its predecessor, GLiNER2.
Small encoder models are cheap but rigid, and large language models are flexible but expensive per document. Fastino released GLiNER2.5 to narrow that gap. The release replaces span enumeration with boundary prediction: the model scores where an entity starts and ends instead of scoring every candidate span against a width grid.
Why this matters
For teams running extraction pipelines at scale, the width ceiling in older GLiNER models was a real constraint, not a theoretical one. Long clauses, multi-sentence entities, anything past the enumerated grid simply got missed or truncated. Boundary prediction sidesteps that by scoring start and end points directly, and the jump to 4,096-word context means documents that used to require chunking, and the stitching logic that comes with it, might not anymore. That matters for anyone building contract review, medical record parsing, or compliance tools where entities don't respect neat token windows.
We're wary of any claim that closes the gap between small encoders and LLMs, since "linear in sequence length" doesn't automatically mean cheap at production volume. But the architectural bet here is sound: span enumeration was always a workaround, not a natural fit for how entities actually appear in text. If Fastino's benchmarks hold up under independent testing, this is worth evaluating for anyone currently burning LLM tokens on extraction tasks that don't need generative reasoning at all.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives UK AI safety test: Rogue AI agent pushed malware into open-source tool
From Archives Hugging Face Reportedly in Acquisition Talks at USD 13 Billion Valuation
Common Questions Answered
How does GLiNER 2.5's boundary prediction approach differ from traditional span enumeration in named entity extraction?
GLiNER 2.5 replaces the traditional span enumeration method with boundary prediction, which scores where an entity starts and ends instead of evaluating every possible span against a maximum width grid. This eliminates the width ceiling that limited entity length in older models and removes the computational scaling problems that occurred as documents grew larger. The new approach allows entities of any length to be properly identified without artificial constraints.
What are the practical advantages of GLiNER 2.5's 4,096-word context window for document processing?
The expanded 4,096-word context window means that documents previously requiring chunking and complex stitching logic can now be processed as complete units, significantly simplifying extraction pipelines. This eliminates the overhead and potential errors associated with reassembling entities that were split across multiple chunks. Teams running extraction at scale can now handle longer documents more efficiently without architectural workarounds.
Why does Fastino position GLiNER 2.5 as a solution between small encoder models and large language models?
Small encoder models are cost-effective but inflexible, while large language models offer flexibility but are expensive to run per document. GLiNER 2.5 aims to narrow this gap by providing a more efficient model that maintains reasonable costs while improving flexibility through its boundary prediction architecture and expanded context capabilities. This makes it a practical middle ground for teams balancing performance and expense in production extraction pipelines.
What specific limitations of older GLiNER models does the boundary prediction method resolve?
Older GLiNER models suffered from a width ceiling that caused long clauses, multi-sentence entities, and content beyond the enumerated grid to be missed or truncated during extraction. The boundary prediction method in GLiNER 2.5 directly scores start and end points rather than relying on a predefined width grid, completely eliminating this constraint. This ensures that entities of any length are properly captured without truncation or omission.
Further Reading
- Introducing GLiNER2.5: Efficient Span-Free Information Extraction - Fastino Labs Blog
- Fastino Releases GLiNER2.5: A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction - MarkTechPost
- Specialized, Open Models | GLiNER, GLiGuard, ... - Fastino Labs
- fastino/gliner2.5-small-v1 - Hugging Face
- An Empirical Study on Finding Spans - ACL Anthology
