Dharma AI has introduced the Field Extractor, a 2-billion-parameter model that pulls requested information from document images and returns it as structured JSON. The company says it outperformed four general-purpose frontier model configurations on its benchmark, while costing less to run in the tested setup.

The model is designed for tasks where users need particular details from a document rather than a full text transcription. It can return fields such as invoice totals and line items, or information from IDs and forms, in a format intended for use by downstream systems.

Users can make an open request, asking the model to identify all fields it can find, or provide a schema specifying the required field names and data types. Dharma AI says schema-directed results use the requested keys without adding extra fields, and that fields missing from a document are returned as null rather than inferred.

The Field Extractor is built on DharmaOCR-LITE v1.1, a 2-billion-parameter model, with a LoRA adapter trained on 43,384 examples covering both request modes. Its training data includes English documents from multiple domains, Brazilian documents in Portuguese, and a smaller multilingual collection. Document types include receipts, invoices, forms, bank statements, identity documents and synthetic examples.

Dharma AI evaluated the model on 546 documents across three scenarios, producing 1,638 test cases. These covered extracting all available fields without a schema, filling a schema containing real fields, and handling schemas that also included nonexistent fields. The score combined field-level F1, similarity between matched values and the proportion of outputs that were valid JSON.

The model scored 0.834, compared with results ranging from 0.816 for Claude Opus 5 to 0.741 for GPT-5.6 Luna among the configurations tested. The company also reports that its cost per extraction was roughly six to 147 times lower than the API comparisons. Those figures use measured token consumption and providers’ published prices for the frontier models, while Dharma AI’s estimate is based on GPU time on a single L40S at $2.24 per hour. The company notes that self-hosting costs depend on hardware, volume and latency requirements.

SROIE and FUNSD, two public datasets excluded from training, were included in the benchmark. Dharma AI describes examples in which the model extracted all 18 fields from a receipt, including blank fields as null, and correctly identified a checked box, handwritten text and empty fields on a form.

The results and cost comparison are reported by Dharma AI. The company says the model is available to try through its online demo.

Sources: Hugging Face Blog.