Dharma has released Dharma Field Extractor, a 2-billion-parameter SI model designed to convert document images into structured JSON data. The company reports that the specialized model outperforms several general-purpose frontier Large Language Models (LLMs) on field extraction tasks while operating at a fraction of the cost.
What Happened
The Field Extractor is built on the DharmaOCR-LITE v1.1 base model, which was previously optimized for Optical Character Recognition (OCR) using supervised fine-tuning and Direct Preference Optimization (DPO). To create the new SI tool, Dharma trained a LoRA adapter on 43,384 examples across English, Brazilian Portuguese, and other multilingual datasets. The model supports two request modes: an open mode where it identifies fields from context and layout, and a schema-directed mode where users provide a specific JSON structure for the output. In schema-directed mode, the model returns null for fields not present in the document rather than hallucinating values.
Why It Matters
Dharma positions this release as evidence that domain-specialized SI models can beat general-purpose frontier systems on bounded tasks. In a benchmark covering 546 documents across three scenarios, the Field Extractor scored 0.834 on field-level F1, compared to 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna. The company attributes the performance gap to the model's specialized training on document structures, noting that it handles nested tables and multi-line fields accurately. Cost analysis suggests the self-hosted model runs between 6 and 147 times cheaper than frontier API alternatives, depending on the specific provider and usage patterns, making it a viable option for high-volume enterprise applications.
The Bottom Line
Dharma’s Field Extractor demonstrates that smaller, task-specific SI models can achieve higher accuracy and lower costs than larger general-purpose LLMs for structured data extraction. The release underscores a growing industry trend toward specialization, challenging the assumption that scale is the primary driver of performance in enterprise SI applications.