Knowledgator Engineering just dropped GLiFormer, a unified encoder framework that handles named-entity recognition, classification, relation extraction, and nested JSON structuring in a single pass. Instead of chaining separate models or relying on slow token-by-token generation from large language models, GLiFormer uses a schema-conditioned approach. You feed it labels and extraction schemas at inference time, and it scores compatibility against shared source encodings. The result is a system that runs on CPU or GPU, installs via pip, and ships with Apache 2.0 licensed weights.
The architecture generalizes the label matching found in GLiNER by introducing ‘anchors.’ These anchors represent the objects each runtime label gets scored against, whether that is a group vector for classification or a record slot for structuring. For nested JSON, the model grounds field values as spans directly from the source text. This design choice is critical because it prevents the model from hallucinating value text that isn’t present in the input. A deterministic decoder then assembles these spans into structured records, handling parent-child links restricted to schema-allowed paths.
Performance numbers are compelling for structured tasks. On a benchmark of 500 examples, GLiFormer Large achieved a 91.10 F1 score for nested JSON structuring, trailing GPT-5.6-luna’s 91.96 by less than a point. More importantly, speed is where this encoder shines. Knowledgator reported a median latency of 69 ms on an NVIDIA RTX PRO 6000 Blackwell GPU for batch size 1. This avoids the overhead of autoregressive generation entirely. However, the model isn’t a universal replacement yet. Relation extraction scores lag behind specialized tools like GLiNER-Relex, averaging just 21.33 micro-F1 for the Large variant compared to 25.6 for its competitor.
With 575.6 million parameters, GLiFormer Large offers a parameter-efficient alternative to massive LLMs for specific extraction pipelines. It beats Gemma-4-E4B on combined NER and classification aggregates while using roughly 14 times fewer parameters. For engineering teams tired of maintaining fragile chains of disparate models, this unified approach offers a pragmatic middle ground. It won’t replace reasoning-heavy LLMs for open-ended tasks, but for high-volume, schema-bound data extraction, it provides a faster, more deterministic path to production.
