Dharma-AI's 2B Field Extractor scores 0.834 on its own benchmark, above Claude Opus 5's 0.816

Dharma-AI has released Dharma Field Extractor, a 2-billion-parameter model that takes a document image and returns structured JSON. According to Dharma-AI, it scored 0.834 on the company's own benchmark, above every frontier configuration it tested, from Claude Opus 5 at 0.816 to GPT-5.6 Luna at 0.741. The announcement was published on Hugging Face on November 4, 2026.

What it does

The model accepts a document image and works in two modes:

• Open: no schema is supplied. The model infers which parts of the page are structured fields (a label followed by a value, a table with a header row) from context and layout, and returns everything it finds as key-value pairs.
• Schema-directed: the user specifies each field with a type (string, number or boolean) under the key names their system already uses. The output follows the schema exactly, with the same keys and no extra fields. A field absent from the document comes back as null rather than an invented value.

Tables are filled row by row, and nested documents return nested JSON. Dharma-AI says one checkpoint handles all document types, with no per-type fine-tuning. Examples include invoices with ordered line items, ID documents, medical requisitions and contracts.

Training

The base is DharmaOCR-LITE v1.1, a new 2-billion-parameter version of the OCR model Dharma-AI released in April, already specialized through supervised fine-tuning and DPO. Dharma-AI trained a LoRA adapter on top of it using 43,384 training examples covering both request modes, in three streams, all annotated at field level:

• English multi-domain documents from the public web
• Real Brazilian documents in Portuguese, sourced internally
• A smaller share of other European and Asian languages

The mix spans receipts, commercial invoices, forms, bank statements, ID documents and synthetic documents.

Benchmark results

The benchmark covers 546 documents across three request scenarios, for 1,638 cases: extraction with no schema, a partial schema naming only real fields, and 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.

Dharma-AI ran the same benchmark on four general-purpose frontier configurations, all with reasoning set to low. Claude Opus 5 used adaptive effort. According to the company:

• The quality gap runs from 1.8 points (Claude Opus 5, the closest) to 9.3 points (GPT-5.6 Luna, the widest).
• Cost runs from roughly 6 times lower (GPT-5.6 Luna) to roughly 147 times lower (GPT-5.6 Sol).

SROIE and FUNSD, two public datasets in the benchmark, were not used in training. On a SROIE receipt, Dharma-AI says the model returned all 18 requested fields correctly, including a multi-line unlabeled address and two blank fields as null. On a FUNSD form with twelve boolean fields, it marked the single checked box true and the other eleven false.

Cost basis

The two sides are priced differently. Frontier costs use measured average input and output tokens per page, reasoning tokens included, against each provider's published list pricing. The Field Extractor is not billed per token: it is self-hosted, and its figure is GPU time on a single L40S at $2.24 per hour. Dharma-AI states this is not a sales price and that self-hosted cost depends on hardware, volume and required latency. Context window and training cutoff are not stated.

What this means

The results are Dharma-AI's own, on its own benchmark, so they show performance on this task design rather than general document understanding. The comparison also sets a self-hosted GPU-hour cost against per-token list pricing, a difference Dharma-AI itself flags. Within those limits, the case is one of task specialization: a 2B adapter-tuned model with schema-locked output and null handling for missing fields targets the step where OCR output still needs custom mapping code before reaching a database. The held-out SROIE and FUNSD examples are the evidence offered that the model generalizes beyond its training data.