Dharma-AI released a 2-billion-parameter model for extracting structured fields from documents. It returns JSON with typed values, matching a supplied schema or inferring fields automatically. The model handles invoices, IDs, medical forms and more without per-type fine-tuning. It outperformed every frontier LLM tested on a 546-document benchmark. Cost per extraction ran 6 to 147 times lower than frontier APIs. Quality scores ranged from 0.834 for the model to 0.741 for the lowest frontier competitor. Training used 43,384 examples across English, Portuguese and other languages. The model is available for trial via Dharma-AI’s portal. It is designed for self-hosted deployment on a single L40S GPU. The benchmark measured field-level F1, value similarity and valid JSON output.
What changed
General-purpose LLMs required careful prompting for field extraction at high cost and latency.
What it unlocks
Extracting structured fields from any document into JSON with a single model, either schema-free or schema-directed.
- 2B parameters
- 43,384 training examples
- 546 benchmark documents
- 1.8 to 9.3 point quality gap vs frontier LLMs
What you need to act on it
- self-hosted deployment
- access to Dharma-AI's portal for trial
Sources