Dharma-AI: a 2B model for document field extraction beats frontier APIs

Dharma-AI, a company that builds small task-specific models, published a post on Hugging Face on 4 November 2026 describing the Dharma Field Extractor, a 2-billion-parameter model that reads a document image and returns only the fields you ask for as structured JSON. The company says it outscored every frontier model it tested on its own benchmark while costing a fraction as much to run.

What it does

The post describes two ways to ask. In open mode you supply no schema, and the model works out which parts of the page are label-and-value pairs or table rows and returns them as key-value pairs. In schema-directed mode you specify each field and its type (string, number or boolean) under the key names your system already uses, and the output follows that schema with the same keys and no extra fields.

Tables are filled row by row. Dharma-AI says that a field requested but absent from the document comes back as null rather than a guessed value. The same checkpoint handles invoices, medical requisitions, contracts and ID documents, with no per-type fine-tuning.

How it was built

The base is DharmaOCR-LITE v1.1, a 2-billion-parameter OCR model the company says was tuned with supervised fine-tuning and DPO. On top of it the team trained a LoRA adapter on 43,384 examples in three streams: English multi-domain documents from the public web, real Brazilian documents in Portuguese sourced internally, and a smaller multilingual share. All are annotated at the field level.

What the benchmark says

Dharma-AI ran 546 documents through three scenarios, for 1,638 cases: extract everything with no schema, fill a partial schema of real fields, and fill a schema that also includes deliberately nonexistent fields. The score is field-level F1, matched by path including table position, multiplied by Levenshtein similarity on matched values and by the share of outputs that are valid JSON.

• Field Extractor (2B): 0.834.
• Claude Opus 5, the closest frontier model, at adaptive effort: 0.816, a gap of 1.8 points.
• GPT-5.6 Luna, the widest gap: 0.741, a gap of 9.3 points.
• All frontier models ran with reasoning set to low, per the post.

On two public datasets not used in training, SROIE receipts and FUNSD forms, the post shows one example each: all 18 fields on a receipt correct, and a twelve-checkbox form read with exactly one box marked true.

The cost claim, and its limits

The company reports a cost per extraction roughly 6 times lower than GPT-5.6 Luna and roughly 147 times lower than GPT-5.6 Sol. The two sides are priced differently. The frontier figures use measured average input, output and reasoning tokens per page against list pricing. Dharma-AI's figure is GPU time on a single L40S at $2.24 an hour, since the model is self-hosted. The post itself says this is not a sales price and that real cost depends on hardware, volume and latency requirements.

The post does not report latency, pages per hour, or how the model fares on layouts unlike the training mix. It also does not say how the model is licensed or where the weights can be downloaded; the post points readers to a demo.

What a team can take from it

If a high-volume job in your business is pulling the same few fields from invoices or forms, a small specialised model is worth a bake-off against the frontier API you use now. Build a sample of a few hundred of your own documents with known answers, score field-level accuracy and null handling, and price both sides on the same basis, including the GPU you would keep running at low volume.