Hugging Face releases Dharma Field Extractor model

On 2026-11-05, Hugging Face announced Dharma Field Extractor on its official blog, describing it as a 2-billion-parameter small model for document structured extraction. Hugging Face says the model extracts exactly the fields you ask for from any document type and directly outputs them as structured JSON. In its testing it outperformed every frontier LLM it tried, at a fraction of the cost.

The release builds on the line of work Hugging Face introduced in April 2026 with DharmaOCR-LITE, a model specialized for OCR across a wide range of documents, from contracts and court filings to invoices, IDs and forms. The paper behind DharmaOCR-LITE made a narrow, well-measured case for domain specialization: a smaller model trained closer to its task can beat general-purpose frontier APIs on that task. Hugging Face has framed OCR as the right tool when what you need is the text on a page, faithful to the original formatting. That fits jobs where the whole document matters, such as a contract you will read end to end or a manuscript you are digitizing. Dharma Field Extractor targets a different job: instead of returning the page, it returns the specific fields you asked for.