Dharma Field Extractor: 2B-Parameter Model Turns Any Document into Structured JSON

Dharma-AI has released Dharma Field Extractor, a 2-billion-parameter vision-language model purpose-built to pull structured fields from document images and return typed JSON — no per-document fine-tuning required. Unlike general-purpose OCR that dumps raw text, or frontier LLMs that need careful prompting and incur high API costs, this model accepts either an open-ended request or a strict schema (field names, types, nesting) and delivers exactly the keys your downstream systems expect.

Two Extraction Modes, One Checkpoint

In open mode, the model infers structure from layout and context — labels, tables, key-value pairs — and returns everything it finds. In schema-directed mode, you declare the exact shape: field names, types (string, number, boolean), and nesting. The same weights handle invoices, medical requisitions, identity documents, and contracts without retraining.

Early benchmarks show the extractor beating much larger frontier models on field-level accuracy while running on a single GPU. For teams processing thousands of documents daily, that translates to lower inference cost, predictable latency, and simpler pipelines — no prompt engineering, no post-processing regexes.

Why it matters for GPU / AI infrastructure: Specialized small models like this change the economics of document intelligence. Instead of routing every page through a massive LLM API, enterprises can deploy a 2B model on modest GPU fleets — even on-prem — keeping data private and costs linear. AiGpu customers running inference workloads on our cloud can spin up dedicated instances for this class of model and scale horizontally as volume grows.