Fastino's GLiNER 2.5 Extracts Entities and Relations Without Repeated LLM Calls

Fastino’s model card lists more than 230,000 downloads for gliner2.5-multi-v1. The 287-million-parameter checkpoint uses an mDeBERTa-v3-base encoder to extract entities, attributes, classifications, and relations across multiple languages. Its schema-driven API targets workloads that otherwise require several specialized encoders or repeated large language model calls.

Boundaries remove the span-width ceiling

Earlier GLiNER releases enumerated candidate spans up to a configured maximum width. That approach increased the number of candidates and prevented the model from returning mentions longer than the limit. GLiNER 2.5 predicts start and end boundaries directly. Fastino describes the resulting inference complexity as linear in document length, although total latency still depends on the schema, label count, and hardware.

Within the reported 4,096-word input window, the model can treat a long product description, legal clause, or multi-sentence quotation as one entity. Boundary prediction removes the architecture’s fixed span-width cap, while the encoded context window still limits the maximum usable span.

One schema coordinates the outputs

The schema builder can request entities, entity-level attributes, document classification, and typed relations in one extract call. Combining those outputs reduces the orchestration required for pipelines that currently pass text through separate named entity recognition, classification, and relation models.

schema = (
    model.create_schema()
    .entities(["person", "organization", "product"])
    .entity_attributes({
        "sentiment": AttributeGroup(
            ["positive", "negative", "neutral"],
            applies_to=["person"]
        )
    })
    .classification(
        "topic",
        ["technology", "business", "sports"]
    )
    .relations(["works_for", "announced"])
)

result = model.extract(text, schema)