Dharma Releases 2B-Parameter Field Extractor That Turns Documents Into Structured JSON

The model is built on DharmaOCR-LITE v1.1, an updated version of the OCR model the company released in April. Dharma added a LoRA adapter trained on 43,384 examples drawn from three sources: English documents from many domains collected from the public web, Brazilian documents in Portuguese sourced internally, and a smaller multilingual set covering other European and Asian languages.

Two Ways to Ask

Users can send a document with no schema, and the model returns every label-and-value pair or table it finds. Or they can supply a schema that names each field and its type, and the output then follows that schema exactly. Dharma says that when a requested field does not appear in the document, the model returns null instead of making up a value, a common weakness when general-purpose language models are used for extraction.

The company argues that OCR alone leaves engineers to write code that maps each string to the right database field, while prompting frontier models to do that mapping becomes expensive at the volumes that invoice processing, medical scheduling or compliance work require.

Benchmark Claims

On Dharma's own benchmark of 546 documents across three request scenarios, 1,638 cases in all, the Field Extractor scored 0.834, the company said. Claude Opus 5, the closest of four frontier configurations tested, scored 0.816, and GPT-5.6 Luna, the furthest behind, scored 0.741. The score combines field-level F1, how closely the extracted values match, and the share of outputs that are valid JSON.

Dharma estimates the model is roughly 6 times cheaper than GPT-5.6 Luna and roughly 147 times cheaper than GPT-5.6 Sol. Frontier costs were calculated from measured token use at list prices with reasoning set to low. Dharma's cost was based on self-hosted GPU time on a single Nvidia L40S at $2.24 an hour, and the company notes that this is not a sales price and that real costs depend on hardware, volume and latency needs.

The results come from the company and have not been verified independently. The post does not say what licence the Field Extractor is released under. The model can be tried through a web portal that requires a login.

For enterprises that process large numbers of documents, the release is another example of small, specialised models competing with frontier APIs on narrow, well-defined tasks where cost and predictable output matter more than general ability.