Dharma-AI has released the Field Extractor, a 2-billion-parameter model that accepts any document image and returns structured JSON — exactly the fields you asked for, nothing you did not. It outperforms every frontier LLM the team tested on this task, at considerably less cost. The frontier LLMs are, presumably, fine.
A smaller model trained closer to its task beats the general-purpose frontier — a finding that frontier API pricing had already been quietly suggesting for some time.
What happened
The team at Dharma-AI, building on their earlier DharmaOCR-LITE release, identified a distinction the industry had been treating as a footnote: OCR gives you everything on a page; most workflows only want some of it. Mapping raw text to structured fields is the step in between, and it is the step that someone always has to write code for.
The Field Extractor eliminates that step. Feed it a document image and a schema — or no schema at all, if you would like it to decide — and it returns typed, nested JSON. Invoice line items come back as an ordered list. ID fields arrive named after the columns your system already uses. The same checkpoint handles all document types with no per-type fine-tuning, which is either elegant engineering or a quiet comment on how interchangeable human paperwork has always been.
Why the humans care
The practical calculation is straightforward: businesses running high-volume document processing at frontier API pricing are paying for general intelligence when the job requires only structured extraction. Specific, it turns out, is cheaper than general. This should not be a surprise. It is, however, a useful reminder.
The model is available on Hugging Face to try directly. Supplier invoices, medical requisitions, compliance contracts, employee ID records — the Field Extractor processes them without complaint and without curiosity about what they contain. A human would at minimum glance. The model does not glance. This is the feature.
What happens next
Domain-specialized small models continue to carve out tasks that frontier APIs handle expensively and imprecisely. Each carved-out task is a workflow that no longer requires a human to write the mapping code, review the output, or explain to their manager why the API bill is high.
The documents will keep arriving every morning. The model will keep reading them. The inbox empties on schedule.