Dharma's 2B Field Extractor turns documents into JSON for $2.24 an hour

Dharma AI says it has built a 2-billion-parameter model that reads a document image and returns structured JSON, and that on its own test it edges past Claude Opus 5 at a fraction of the running cost. The claims come from a Hugging Face blog post by Dharma-AI, a vendor writing about its own product, and every figure below is Dharma's.

What it does

The model works in two modes. In open mode it extracts every field it finds. In schema-directed mode it fills a schema you supply, with string, number and boolean types plus nested structures and tables. Dharma says a field that is missing from the document comes back as null rather than an invented value.

It is a LoRA adapter trained on 43,384 examples over DharmaOCR-LITE v1.1, a 2B checkpoint that had already been through supervised fine-tuning and DPO. Training data covers English web documents, Brazilian Portuguese documents the company sourced itself, and a smaller share of other European and Asian languages: receipts, invoices, forms, bank statements, ID documents and synthetic pages.

The benchmark

Dharma ran the test itself: 546 documents and 1,638 cases across three scenarios, covering no schema, a partial schema of real fields, and a schema that includes fields the document does not contain. The score is field-level F1, multiplied by the Levenshtein similarity of matched values and by the share of valid JSON outputs.

• Dharma Field Extractor: 0.834.
• Claude Opus 5: 0.816, run with adaptive effort.
• GPT-5.6 Luna: 0.741. The post says frontier models ran with reasoning set to low.

On two public sets held out from training, Dharma reports that the model got all 18 receipt fields right on SROIE, returning blanks as null, and read a checkbox form on FUNSD correctly, including one nested group of five sub-options.

The cost claim

Dharma prices the frontier models from measured tokens per page at published list prices and says the extractor is roughly 6 times cheaper than GPT-5.6 Luna and roughly 147 times cheaper than GPT-5.6 Sol. The extractor's own cost is GPU time, one L40S at $2.24 an hour when self-hosted. The post says this is not a sales price and that real cost depends on hardware, volume and latency.

What the post leaves out

• Licence and weights: not stated. The post links a web portal to try it; DharmaOCR-LITE is on Hugging Face, and its page lists that model as 4B while the post describes the base as 2B.
• Independent results: none. The test set, the scoring and the choice of reasoning settings are Dharma's.
• Throughput and latency: not given, so the $2.24 an hour cannot yet be turned into a cost per page.
• Other limits: the post lists none.

For a team processing invoices or receipts, the useful part is the shape of the claim: a small, self-hosted model matching a frontier model on a narrow, well-defined extraction job. That is plausible, and it is also exactly what a vendor would test. Run it on a sample of your own documents, including ones with missing fields, before drawing a cost comparison.