Brazilian AI firm Dharma releases 2B-parameter document field extractor, claims self-tested costs far below frontier LLMs
Dharma-AI has launched a 2-billion-parameter small language model designed specifically for extracting structured fields from documents. Vendor self-tests indicate it surpasses frontier LLMs in quality while significantly reducing costs, offering a cheaper alternative for high-volume document processing.
Brazilian artificial intelligence startup Dharma-AI released Dharma Field Extractor, a 2-billion-parameter small language model (SLM), on November 4, 2026. The model is designed to directly extract specific fields requested by users from various types of document images and output them as structured JSON. Unlike traditional Optical Character Recognition (OCR) technology, which only returns all text printed on a page, Field Extractor understands context and layout to map specific information—such as supplier names, total amounts, or patient names—from invoices, medical requisitions, or ID cards into designated key-value pairs.
Technically, the model is built upon DharmaOCR-LITE v1.1, trained through supervised fine-tuning and Direct Preference Optimization (DPO), with an additional LoRA adapter trained on a dataset of 43,384 examples. The training data covers English multi-domain documents, Portuguese Brazilian local documents, and a smaller share of other European and Asian languages. The model supports two request modes: an "open" mode where the model identifies and extracts all structured fields on its own, and a "schema-directed" mode where users specify exact field names and data types, with the model returning null for fields not present on the document.
According to internal benchmark results published by Dharma-AI, the model achieved a composite score of 0.834 across 546 test documents, outperforming several frontier general-purpose LLM configurations, including Claude Opus 5 and GPT-5.6 Luna. In terms of cost, because the model is designed for self-hosted deployment, its inference cost on a single L40S GPU is approximately 6 to 147 times lower than calling frontier APIs. This result highlights the potential economic advantage of domain-specialized small models in high-frequency enterprise applications. It should be noted that these performance comparisons and cost savings are based on benchmarks and hardware environments constructed by Dharma-AI itself, and have not yet been independently reproduced by third parties.