Dharma-AI, a Brazilian research lab specializing in Small Language Models (SLMs), has announced the release of its ‘Dharma Field Extractor.’ This new model is designed to efficiently extract structured data from any document type, converting it into JSON format. The Field Extractor aims to provide a cost-effective and high-performing alternative to general-purpose Large Language Models (LLMs) for tasks involving document data extraction.
The Dharma Field Extractor operates in two modes: ‘Open,’ where it identifies and extracts all structured fields it detects in a document, and ‘Schema-directed,’ where users specify the exact fields and data types they require. This flexibility allows the model to adapt to various use cases, from extracting line items on invoices to capturing patient details from medical forms or processing information from identification documents.
According to details received by The Chenab Times, Dharma-AI positions the Field Extractor as a solution to the limitations and costs associated with using large, general-purpose LLMs for high-volume, specific data extraction tasks. The company highlighted that while Optical Character Recognition (OCR) provides raw text from documents, it often lacks the context to distinguish between different data fields. General-purpose LLMs can perform this task but at a higher operational cost and latency, especially when deployed at scale.
Performance and Cost Efficiency
Dharma-AI claims that their 2-billion-parameter Field Extractor model outperforms frontier LLMs in field extraction tasks, achieving a score of 0.834 on a benchmark where Claude Opus 5 scored 0.816 and GPT-5.6 Luna achieved 0.741. This higher quality is reportedly delivered at a significantly lower cost per extraction. The model is designed to be self-hosted, with operational costs dependent on hardware, volume, and latency requirements. Dharma-AI provided a comparison indicating that their model can be 6 to 147 times less expensive to run than frontier API services.
The training of the Dharma Field Extractor involved a base model, DharmaOCR-LITE v1.1, which was further fine-tuned using a dataset of 43,384 examples. This dataset included diverse English documents from the public web and specific Brazilian documents in Portuguese, covering both extraction modes.
Dharma-AI emphasizes its philosophy that specialized models, trained for specific tasks, can outperform larger, more general models. This approach aims to provide businesses with AI solutions that are not only powerful but also economically viable and efficient for their particular needs. The company’s research also focuses on mechanistic interpretability of SLMs and GPU utilization optimization to further enhance the performance and efficiency of their models.
❤️ Support Independent Journalism
Your contribution keeps our reporting free, fearless, and accessible to everyone.
Or make a one-time donation
Secure via Razorpay • 12 monthly payments • Cancel anytime before next cycle
(We don't allow anyone to copy content. For Copyright or Use of Content related questions, visit here.)
The Chenab Times News Desk
