Fastino releases GLiNER2.5 extraction models
Fastino has released GLiNER2.5, which swaps span enumeration for boundary prediction so extraction scales linearly with document length, handles entities of any length and returns schema-valid knowledge graphs.
Fastino announced GLiNER2.5, calling it the biggest upgrade yet to its GLiNER family of compact information-extraction models. Three variants are on Hugging Face under Apache 2.0: gliner2.5-base-v1 (0.2B parameters), gliner2.5-multi-v1 (0.3B) and gliner2.5-small-v1 (74M).
Fastino says two changes drive the release. First, the architecture scores where an entity starts and ends rather than enumerating candidate spans, so inference scales linearly with document length. Second, the training data was generated with the company's Fastino Data Agent instead of being assembled from existing datasets.
GLiNER2 only enumerated spans up to a fixed width, typically about twelve words, so it never scored longer entities. GLiNER2.5 has no width limit. According to the post, a forty-word contract clause costs the same to locate as a two-word name. The lower memory use also allowed training on sequences of up to 4,096 words. The library now has built-in chunking for longer inputs, which maps results back to character offsets in the original text and merges duplicates where chunks overlap.
The model also extracts entities and relations jointly in one forward pass. A beam search then checks each candidate against the user's schema rules as it builds the result, so the output graph has no relations pointing to missing entities, no cardinality violations and no cycles. Other additions include constrained classification and span attributes.
Fastino says GLiNER2.5 beats GLiNER2 on average F1 across 16 classification and extraction benchmarks, including a 24.75-point gain on XNLI. The excerpt does not include per-benchmark results or any comparison with models other than GLiNER2.