What is Dharma Field Extractor?
Dharma Field Extractor is a document field-extraction model from Dharma-AI S.A., a Rio de Janeiro company that builds small specialised models for enterprises. Announced on the Hugging Face blog on November 4, 2026, it is a LoRA adapter trained on top of DharmaOCR-LITE v1.1, a new 2-billion-parameter version of the company's OCR model. Neither the adapter nor the v1.1 base is published on Hugging Face; the only public Dharma model is the earlier DharmaOCR-LITE, under a non-commercial licence.
It takes a document image and returns JSON in one of two modes. In open mode it extracts every field it finds, nesting them as the document does, including table rows such as invoice line items. In schema mode you supply the keys and a type for each (string, number or boolean) and get back exactly those keys, with null for anything not on the page. The adapter was trained on 43,384 examples: English documents from the public web, real Brazilian Portuguese documents from Dharma-AI's own sources, and a smaller set in other European and Asian languages, covering receipts, invoices, forms, bank statements and IDs.
Is Dharma Field Extractor worth it?
The headline result is narrow and self-reported. On Dharma-AI's own benchmark of 546 documents, each tested three ways (no schema, a schema of real fields, and a schema salted with fields that do not exist), it scores 0.834 against 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna. The frontier models ran with reasoning set to low. The test set, the metric and the winner all come from the same company, and the post publishes no training settings or code to reproduce it.
The metric is a sensible one for production work. It multiplies field-level F1, with fields matched by their path including position in a table, by edit-distance similarity on the values and by the share of outputs that parse as JSON. That punishes the failures that break a pipeline: an invented field, a wrong line item, a stray bracket. The schema mode's habit of returning null for absent fields, rather than guessing, is the behaviour accounts-payable teams want.
The cost argument is the real pitch. Self-hosted on one NVIDIA L40S at $2.24 an hour, Dharma-AI puts the run at roughly 6 times cheaper than GPT-5.6 Luna and 147 times cheaper than GPT-5.6 Sol, while adding that $2.24 is not a sales price. A 2B model on one mid-range GPU is a plausible way to clear millions of receipts a month.
What stops a recommendation is access. There are no weights, no API price and no published terms for data handling, only a demo portal and a sales contact. The earlier DharmaOCR-LITE weights are non-commercial and need a separate written licence for any hosted or commercial use, which suggests the same model for this one. Open alternatives such as GLM-OCR and PaddleOCR-VL can be downloaded and tested on your own documents today; this one asks for a conversation first.
Where it earns its place
An accounts-payable lead at a Brazilian distributor handling notas fiscais and supplier invoices
Pull totals, tax IDs and line items into the ERP without retyping
Send each scanned invoice with a schema using the ERP's own field names and post the JSON straight in, flagging nulls for review. The model saw real Brazilian documents in training; deployment and price have to be negotiated with Dharma-AI.
A platform engineer replacing a frontier-model extraction step that costs too much per page
Cut the per-document bill on a high-volume receipt pipeline without losing accuracy
Score a sample of your own documents on the demo against the current model before signing anything, using the same three cases Dharma-AI tests (no schema, real fields, fake fields); the published 6 to 147 times saving assumes self-hosting on an L40S at $2.24 an hour.
Before you commit
- Setup
- Try a document on the demo portal linked from the launch post; production use goes through Dharma-AI's sales team. No public weights, SDK or API reference.
- Integrations
- JSON output keyed to your own field names; Dharma-AI says its models run on-premise or in a private cloud. No third-party integrations are listed.
- Data and privacy
- Dharma-AI says clients keep full control of their data in their own environment; retention, training use and certifications are not published. The public DharmaOCR-LITE weights require a written licence from Dharma-AI S.A. for commercial, hosted or government use.
- Cost at scale
- Not published. Dharma-AI's benchmark costing uses one NVIDIA L40S at $2.24 an hour and claims roughly 6 to 147 times lower cost than GPT-5.6 Luna and GPT-5.6 Sol.
- Limits
- Takes document images; field types limited to string, number and boolean. Generalisation outside the training set is shown on one SROIE receipt and one FUNSD form, not a full benchmark.
What’s good
- Schema mode returns exactly your keys and null for missing fields
- Handles nested fields and table rows such as invoice line items
- Trained on real Brazilian Portuguese documents as well as English
- 2B parameters; Dharma-AI's cost figures assume one L40S
What’s not
- No public weights, API or price
- Benchmark, metric and comparison are all Dharma-AI's own
- Image input only; PDF handling is not described
- Earlier Dharma weights are licensed for non-commercial use only
Best for
Brazilian and Latin American finance and operations teams processing high volumes of receipts, invoices and forms who are willing to buy a private deployment from a small vendor.
Dharma Field Extractor alternatives
Dharma Field Extractor: common questions
- Is Dharma Field Extractor worth it?
- Dharma Field Extractor makes a credible case that a 2B model can match a frontier model at filling in invoice fields, but until you can download it or see a price, the case is all you can buy. We rate it 3.3 out of 5.
- How much does Dharma Field Extractor cost?
- No price published. Commercial use and deployment by contract with Dharma-AI; a demo portal is linked from the launch post.
- What are the downsides of Dharma Field Extractor?
- No public weights, API or price; benchmark, metric and comparison are all Dharma-AI's own; image input only; PDF handling is not described; earlier Dharma weights are licensed for non-commercial use only.
- Who is Dharma Field Extractor best for?
- Brazilian and Latin American finance and operations teams processing high volumes of receipts, invoices and forms who are willing to buy a private deployment from a small vendor.
- How much does Dharma Field Extractor cost for a team or in production?
- Not published. Dharma-AI's benchmark costing uses one NVIDIA L40S at $2.24 an hour and claims roughly 6 to 147 times lower cost than GPT-5.6 Luna and GPT-5.6 Sol.
- Is Dharma Field Extractor safe for company data?
- Dharma-AI says clients keep full control of their data in their own environment; retention, training use and certifications are not published. The public DharmaOCR-LITE weights require a written licence from Dharma-AI S.A. for commercial, hosted or government use.
- What are the best Dharma Field Extractor alternatives?
- The closest alternatives we have reviewed are GLM-OCR, PaddleOCR-VL-1.6, Mistral OCR 4.1 and Chandra OCR 2.
Answers checked Nov 9, 2026.