Knowledgator's 575m model reports 91.10 F1 on JSON extraction

single source· 1 articles · confidence: low · first seen 2026-09-16 21:19 UTC

What this means for you

Nothing to act on yet. The comparison comes from one release post with no evaluation date, no harness and no independent replication, so it is a claim rather than a result. If you extract structured data at volume, a non-generating encoder is worth tracking — but ask for the evaluation setup before switching anything.

Knowledgator has released GLiFormer Large, a 575m-parameter encoder for pulling nested JSON out of documents. An encoder labels spans of the input text rather than generating an answer token by token, and the release says each extracted value is grounded in a span of the source. Knowledgator puts it at 91.10 F1 on the task, against 91.96 for GPT-5.6-luna. No evaluation date, harness or independent replication is given, and the coverage is a single short write-up.

Models in this story

Key facts

  • ·GLiFormer Large is a 575m-parameter encoder released by Knowledgator. source
  • ·The model is reported at 91.10 F1 on nested JSON extraction. source
  • ·GPT-5.6-luna is reported at 91.96 F1 on the same task. source
  • ·The release says every extracted value is grounded in a span of the source text. source
  • ·The model is described as producing no generated tokens. source
  • ·No evaluation date or harness is given for the reported score. source

What the sources say

  • MarkTechPost — Only source: a short release note giving parameter count, F1 score and a comparison model.

Sources

The original reporting. Follow these — they did the work.

← the wire