Knowledgator releases GLiFormer, a 575M encoder extracting nested JSON without token generation

Knowledgator has released GLiFormer, a single encoder model that handles named-entity recognition, text classification, relation extraction, nested JSON construction, and embeddings in one forward pass. The company published two checkpoints on Hugging Face: Base v1 at 264.2 million parameters and Large v1 at 575.6 million.

Instead of chaining multiple models in a pipeline, the architecture encodes the input text once and then runs different schemas as local task heads over the same shared representation. The mechanism relies on an "anchor" that represents each label at runtime — a class vector for classification, an entity pair for relations, or a record slot for structure — allowing schemas to be swapped at inference without retraining.

JSON construction proceeds in four stages: anchoring field values as spans taken directly from the source text, assigning spans to record slots via Hungarian matching, predicting allowed parent-child relationships per schema, and deterministic decoding to the final structure. Because values are always spans from the input, the model cannot hallucinate text that does not exist in the source, though span selection, record assignment, and hierarchy can still be wrong. Both checkpoints use a DeBERTa backbone that received additional pre-training on 100 billion tokens, with five task heads, a maximum span width of 12 tokens, and 100 record anchors.

On 500 nested-JSON examples, Knowledgator reports 91.10 F1 for Large and 87.20 for Base, compared with 91.96 for GPT-5.6-luna and 82.56 for GPT-5-mini, measured on an order- and boundary-tolerant metric rather than exact match. Across 13 classification datasets, Large reaches an average macro-F1 of 75.03, still below GPT-5-mini's 79.79. On CrossNER's five domains, Base scores 65.10 and Large 64.35 versus 70.74 for Gemma-4-31B-IT. The gap widens in relation extraction: 21.33 and 18.94 micro-F1 respectively, against 25.6 for GLiNER-Relex and 25.08 for Gemma. The report notes that Large outperforms Gemma-4-E4B on combined NER and classification with 14 times fewer parameters, but all figures come from the vendor without external validation.

The speed claim rests on an analytical estimate, not a head-to-head run against an LLM. Knowledgator measured Base on 40 schema documents at batch size 1: a median of 69 milliseconds on an RTX PRO 6000 Blackwell GPU in FP16, and 547 milliseconds on an 8-thread EPYC 9B45 CPU in FP32. The "up to 95.8x faster" figure assumes a prefill rate of 2,000 tokens per second and a generation rate of 60 tokens per second, excluding queues, network latency, hidden reasoning, and without assuming accuracy parity. In other words, it is a theoretical calculation under ideal conditions, not a real comparative benchmark.

Both models are available now under the Apache 2.0 license, installable via pip install gliformer, and run on CPU or GPU without cloud dependencies. Training comprised 1,357,671 broad multi-task examples and 372,090 focused post-training examples. The code and model cards on Hugging Face include usage examples, but in the absence of independent evaluation, users should verify fit for their task before deploying to production.