Cover: AI-generated editorial illustration.
On November 4, 2026, Brazilian AI lab Dharma-AI published Dharma Field Extractor, a 2-billion-parameter model that takes a document image and returns structured JSON: typed fields, nested tables, nulls where data is absent.
What it does
Two modes: open (no schema supplied, the model infers fields from layout) and schema-directed (you provide exact key names and types, and the output matches them precisely). Ask for a field not present in the document and you get null, not a hallucinated value. The same checkpoint handles invoices, IDs, medical forms, and bank statements without per-document fine-tuning.
Benchmark results
Dharma-AI ran a 546-document benchmark scoring field-level F1 combined with Levenshtein value similarity and valid-JSON rate. According to their published results, the Field Extractor scored 0.834, above Claude Opus 5 (0.816) and GPT-5.6 Luna (0.741). Cost per extraction ran 6x to 147x lower than those frontier APIs, based on measured token counts against published list pricing. Their model runs self-hosted on a single L40S at $2.24/hr GPU time; they note this is not a sales price and that self-hosted cost depends on your hardware and volume.
Why this matters in practice
Document field extraction is a high-volume, repetitive task: supplier invoices every morning, patient intake forms, ID onboarding. At frontier API pricing and latency, this gets expensive fast. The practical takeaway is that for a bounded, well-defined extraction task, a small specialist model is worth evaluating seriously before defaulting to a general-purpose API.
The model is built on DharmaOCR-LITE v1.1 with a LoRA adapter trained on 43,384 examples across English, Brazilian Portuguese, and several other languages. SROIE and FUNSD test sets were held out from training.
If you are building document processing pipelines, this is worth a look alongside general-purpose vision models available through OpenNan.