A 2 billion parameter model turns documents into JSON fields

The team behind DharmaOCR released Field Extractor, a model with 2 billion parameters that reads an image of a document. It gives the fields as JSON, a data format that other software reads, and gives null for a field that is not in the document. A user can tell it to find all fields, or give a schema, an accurate list of fields and types.

How it works

The model comes from DharmaOCR-LITE v1.1, a model that reads text in images. The team added a LoRA adapter, a small added layer, with training data in English, Portuguese and other languages. One checkpoint, a copy of the model, operates for all types of document.

The numbers

• Score difference: 1.8 points to Claude Opus 5, 9.3 points to GPT-5.6 Luna
• Cost: 6 times lower than GPT-5.6 Luna, 147 times lower than GPT-5.6 Sol
• Cost of the new model: one L40S GPU at $2.24 an hour

What is not known

The team tells that its cost is not a sales price. The frontier models have list prices, but the new model has GPU time. That cost changes with the hardware, the volume and the speed that a deployment must have.

Insight: Small models get narrow tasks from frontier APIs

The team compares the GPU cost of its model with the list prices of frontier models. Baidu has an open model that reads text in images, and Cohere has a reranker that cuts the text a large model must read. This reads as a division of the work, in which small models do high volume steps and frontier models keep the hard steps. Teams with many documents must find the price of each step.