GLiNER2.5

GLiNER2.5 is a model from Fastino that pulls out specific pieces of information—names, dates, clauses—from documents. Older methods worked like scanning a sentence with a highlighter, drawing every possible 'this could be the span' boundary and checking each one against the rest. But the longer a candidate span got, the more computation it took, so these methods usually capped spans at around twelve words. A long forty-word clause, for instance, would simply never have been considered a candidate in the first place.

GLiNER2.5 changes the approach. Instead of drawing out every possible span in advance, it only marks 'this looks like a start point' and 'this looks like an end point.' Think of it like underlining a passage by marking just the start and end dots and connecting them, rather than shading in the whole thing from start to finish. Because of this, the cost barely changes no matter how long the span is, effectively removing the upper limit on entity length.

Fastino released the model in three sizes: 74M, 194M, and 287M parameters. Unlike a general-purpose chatbot that can answer any question, it's a tool specialized for a single task—pulling information out of documents.

GLiNER2.5 comes up in articles in the context of 'removing the length limit on entity extraction.' One easy point to misunderstand: this isn't a general-purpose model like a conversational chatbot, but one specialized for a single task—information extraction. Its parameter count also ranges from 74M to 287M, far smaller than what's typically called a large language model.