New 2B-Parameter AI Outperforms Frontier Models at Document Field Extraction

A new specialized artificial intelligence model called Dharma Field Extractor can take any document image and return structured data in JSON format. The system allows users to either ask for all available data or provide a strict schema that the output must follow. It handles various document types, including invoices, medical forms, receipts, and identification cards, without requiring specific fine-tuning for each format.

Beating Larger Competitors at Lower Costs

The model contains two billion parameters, which are the variables an AI adjusts during training to learn patterns. Developers tested the system against several large, general-purpose frontier language models on a benchmark covering 546 documents. The compact model scored higher on accuracy while running on a single graphics processing unit. Operating costs for self-hosting the model were significantly lower than the per-token pricing of the tested cloud application programming interfaces.

Targeted Training Over Raw Scale

The creators built the new software on top of an existing optical character recognition checkpoint. They added specialized training using tens of thousands of examples to ensure the system returns accurate text fields and nested lists. This result supports the idea that smaller models, when carefully trained for specific bounded tasks, can surpass massive general-purpose systems while using fewer computing resources.