Dharma-AI's 2B Field Extractor turns documents into typed JSON and beats frontier models on its own benchmark

Plain OCR gives you all the text on a page and leaves someone to write code mapping each string to a database field[2]. Field Extractor skips that step. In open mode you send no schema, and the model works out from layout and context which parts of the page are fields, then returns them as key-value pairs[1]. In schema mode you list the fields you want, each typed as a string, number or boolean, under the key names your system already uses, and the output matches that shape exactly with no extra keys[1]. A field that isn't on the document comes back as null rather than a guess, and tables such as invoice line items come back as ordered lists[1].

The base is DharmaOCR-LITE v1.1, a new 2B version of the OCR model Dharma released in April, already tuned with supervised fine-tuning and DPO[1]. On top of it the team trained a LoRA adapter on 43,384 examples[1]. The data mixes English documents from the public web, real Brazilian documents in Portuguese, and a smaller share in other European and Asian languages, covering receipts, invoices, forms, bank statements and IDs[2]. One checkpoint handles every document type with no per-type fine-tuning[1].

Dharma's benchmark has 546 documents and 1,638 cases across three scenarios: all fields with no schema, a partial schema, and a schema that includes fields that don't exist on the page[2]. Field Extractor scored 0.834, against 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna[2]. Dharma puts its cost at roughly 6 to 147 times lower than the frontier runs, but the comparison mixes pricing methods: frontier costs use measured tokens at list prices, while Dharma's figure is GPU time on one rented L40S at $2.24 an hour[1]. All of these numbers come from Dharma itself[2].

The announcement links a demo for trying the model on your own documents, but doesn't state licence terms or confirm downloadable weights for Field Extractor[2]. The base Dharma-OCR-LITE model family is published on Hugging Face[3].

Pulling a handful of fields off invoices, IDs and contracts is one of the most common paid uses of large language models, and it rarely needs a frontier model. A 2B model that returns schema-exact JSON and refuses to invent missing values is the kind of narrow specialist that makes that work cheaper and easier to trust, once independent tests back up the claims.