Dharma-AI released a 2-billion-parameter Field Extractor model on Hugging Face. The model is designed to pull structured JSON from various document types, including invoices and IDs. It was trained on 43,384 examples covering receipts, forms, and identification cards.
The model scored 0.834 on a benchmark containing 546 documents across three extraction scenarios. This performance beat the scores of Claude Opus 5 (0.816) and GPT-5.6 Luna (0.741). It runs on a single L40S GPU at roughly $2.24 per hour.
The base checkpoint is DharmaOCR-LITE v1.1, which was optimized through supervised fine-tuning and DPO. Dharma-AI previously showed a 3-billion-parameter model beating frontier APIs on Brazilian Portuguese OCR. This new tool offers a cheaper alternative for developers needing document processing.
