01 / The short version
What happened
Dharma Field Extractor is a 2-billion-parameter model built on DharmaOCR-LITE v1.1 that extracts specific fields from documents into structured JSON. It outperforms frontier LLMs like Claude Opus 5 and GPT-5.6 Luna on a 546-document benchmark, achieving higher F1 scores at significantly lower cost. The model supports open requests or schema-guided extraction across diverse document types including invoices, IDs, and forms, returning typed JSON with nulls for missing fields. It was trained on 43,384 field-level annotated examples and runs self-hosted on a single L40S GPU.
See the exact references02 / Key takeaways
What you need to know
- 01
Field Extractor outperforms frontier LLMs by 1.8 to 9.3 F1 points on document field extraction
- 02
Cost is 6 to 147 times lower than frontier APIs when self-hosted on a single L40S GPU
- 03
Model handles multi-domain documents with schema-agnostic and schema-constrained extraction modes
Keep in perspective
What to watch for
Benchmark results are publisher-reported; independent validation of F1 scores and cost comparisons against other specialized extraction models is not provided.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/Dharma-AI/dharma-field-extractorAI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works