Dharma-AI published details of Field Extractor on Hugging Face on November 4, 2026. It is a 2-billion-parameter model for pulling specific fields out of documents. Plain OCR returns all the text on a page and leaves someone to write code that maps each string to a database field. This model returns structured JSON instead.
Two ways to ask
The model accepts a document image plus one of two request types:
- Open: no schema is supplied. The model works out from layout and context which parts are fields, such as a label followed by a value or a table with a header row, and returns key-value pairs.
- Schema-directed: you specify the fields you need, assigning each a type (string, number, or boolean), using the key names already in place in your system. The output follows that shape exactly, with no extra keys. A field that is not on the document comes back as null instead of an invented value.
Tables are filled row by row. An invoice, for example, returns its line items as an ordered list, each with quantity and value, next to the totals. According to Dharma-AI, one checkpoint handles every document type with no per-type fine-tuning.
How it was built
The base is DharmaOCR-LITE v1.1, a new 2B version of the company's OCR model released in April, already tuned with supervised fine-tuning and DPO. On top of it the team trained a LoRA adapter on 43,384 examples covering both request modes. The data has three streams: English documents collected from the public web, real Brazilian documents in Portuguese sourced internally, and a smaller multilingual share in other European and Asian languages. All were annotated at field level and cover receipts, invoices, forms, bank statements and ID documents, among others.
What the benchmark says
The benchmark has 546 documents and 1,638 cases across three scenarios: all fields with no schema, a partial schema of real fields, and a schema that also includes fields that do not exist. The score combines field-level F1, value similarity on matched fields and the share of valid JSON outputs.
Field Extractor scored 0.834. The four frontier configurations ranged from 0.816 (Claude Opus 5) to 0.741 (GPT-5.6 Luna). Dharma-AI puts its cost at roughly 6 to 147 times lower, depending on the comparison model.
The two sides are priced differently. Frontier costs are based on measured tokens per page at list prices, with reasoning set to low. The $2.24-per-hour rate Dharma-AI cites is for renting one L40S GPU for an hour. The company says this is not a sales price and that self-hosted cost varies with hardware, volume and latency. Dharma-AI also reports tests on two public datasets, SROIE and FUNSD, that were held out of training. All of these numbers come from the vendor itself, and no independent replication is cited.
What is available
The post links a demo for trying the model on your own documents. The source does not state licensing terms, weight availability or a commercial offering for Field Extractor, so those remain unconfirmed.