Knowledgator Releases GLiFormer: A 575M-Parameter Encoder That Hits 91.10 F1 on Nested JSON Extraction With…

What changed

Knowledgator launched GLiFormer Large, a 575 million parameter encoder designed specifically for extracting nested JSON data. Unlike many language models, it operates without generating tokens as output, instead directly grounding extracted values in their original source spans. On the standard nested JSON extraction task, GLiFormer scored 91.10 F1, closely trailing GPT-5.6-luna’s 91.96.

Why builders should care

Parsing nested JSON reliably is a known challenge, especially when precision matters for data extraction pipelines. GLiFormer’s approach bypasses token generation in favor of using source span grounding, which reduces output complexity and potential error propagation. This can lead to cleaner, more trustworthy extraction results for applications such as document understanding, knowledge base population, or any workflow requiring structured data pulled directly from text.

GLiFormer’s competitive accuracy at a relatively modest model size (575M parameters) suggests it offers a more efficient alternative to massive LLMs, lowering compute costs without sacrificing performance on this task. Builders working with complex structured data extraction now have a practical tool that cuts through some of the unreliability of token-generating models.

The practical takeaway

Operations dependent on nested JSON extraction can embed GLiFormer to speed up workflows while maintaining near state-of-the-art quality. The grounding mechanism means fewer manual checks and less noise from model hallucinations in the output. This is key for sectors like finance, legal, or healthcare, where data accuracy is critical.

Lower inference costs and higher precision shift the balance towards encoder-centric solutions that handle structured outputs internally rather than through generated text. For AI teams, this opens up more scalable deployments and potentially faster iteration cycles using smaller models that don’t compromise much on quality.

What to watch next

It will be important to see if Knowledgator extends GLiFormer’s architecture to more tasks or larger models, possibly competing with or integrating into bigger LLM frameworks. Watch for adoption signals in industries requiring structured data extraction at scale.

Also, monitor how other developers respond in the encoder-only space, as GLiFormer puts pressure on token-centric generation models to address reliability and efficiency challenges, particularly in enterprise contexts with strict precision demands.

AI Quick Briefs Editorial Desk