Pulling the total, the date and the line items out of a scanned invoice is dull, high-volume work, and many teams now hand it to a frontier model through an API and pay per page. Dharma AI, a company that builds small specialised models, argues that a 2-billion-parameter model trained for this one job does it better and far more cheaply. If that holds on real paperwork, it changes the sums for anyone processing receipts, forms or bank statements at scale.
Field Extractor reads a document image and returns typed JSON. Without a schema, it finds labels, values and tables on its own; with a schema, it returns exactly the requested keys, each as a string, number or boolean, and a null where a field is absent rather than a guess. Nested data such as invoice line items comes back as ordered lists. It is built on the company's DharmaOCR-LITE v1.1 with a LoRA adapter trained on 43,384 examples of receipts, invoices, forms, statements and ID documents, mostly in English and Brazilian Portuguese.
On Dharma's benchmark of 546 documents, each tested three ways for 1,638 cases, it scored 0.834 against 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna. One of the three scenarios deliberately asks for fields that do not exist, which tests whether a model invents values, a common failure in extraction work. On cost, Dharma says it ran about six times cheaper than GPT-5.6 Luna and about 147 times cheaper than GPT-5.6 Sol, priced on one L40S GPU at $2.24 an hour.
Our Dharma Field Extractor review notes that there are no public weights or price yet, and the blog post does not change that: the only access mentioned is a login portal. That matters for the comparison. A self-hosted 2B model is attractive precisely because you can run it on your own hardware, and until weights or an API price appear, the cost advantage is a calculation rather than an offer. Teams already using Claude Opus 5.5 or GPT-6 Luna for extraction have no immediate reason to move.
The doubts are the usual ones for a vendor benchmark. Dharma built the test, chose the scoring formula and ran the rivals with reasoning set to low, which may understate them. Its costs are GPU time, while the frontier costs are list prices, and Dharma itself says the hourly figure is not a sales price. Only two datasets were held out from training, and the showcase examples are single documents. The margin over Claude Opus 5 is under two points. Testing on a sample of your own documents is the only way to know whether the gap is real for your work.