Dharma-AI's 2B Field Extractor tops frontier LLMs on its own benchmark

Dharma-AI says a 2-billion-parameter model it fine-tuned to pull named fields out of document images outscored every frontier model it tested on its own benchmark, at a fraction of the cost. The company published the claim on 4 November 2026 on the Hugging Face blog. The model is called Field Extractor.

What it does

According to the post, the model takes a document image and returns typed JSON. In open mode it finds label-and-value pairs and tables on its own. In schema mode you list the fields and their types (string, number or boolean), and the output uses exactly those keys with no extras. Dharma-AI says it returns null for a field the document does not contain, rather than inventing a value.

It is built on DharmaOCR-LITE v1.1, the company's 2B OCR model, with a LoRA adapter trained on 43,384 examples. The authors say the data mixes English multi-domain documents from the public web, real Brazilian documents in Portuguese sourced internally, and a smaller multilingual share.

The numbers, and who ran them

Dharma-AI's benchmark covers 546 documents in three request scenarios, 1,638 cases in all: every field with no schema, a partial schema, and a schema that includes deliberately nonexistent fields. The score multiplies field-level F1 by the Levenshtein similarity of matched values and by the share of outputs that are valid JSON.

On that score the company reports 0.834 for its model, against 0.816 for Claude Opus 5 and 0.741 for GPT-5.6 Luna. It lists the gap as 1.8 points to 9.3 points across four frontier configurations, all run with reasoning set to low. Two public datasets, SROIE and FUNSD, were held out from training, and the post shows one example from each.

The cost comparison

The post reports cost 6 times lower than GPT-5.6 Luna and about 147 times lower than GPT-5.6 Sol. The two sides are measured differently. Frontier figures use measured tokens per page, reasoning included, at published list prices. Dharma-AI's figure is GPU time on a single L40S at $2.24 per hour. The authors say this is not a sales price and that self-hosted cost depends on hardware, volume and latency.

What the post leaves out

• Whether the benchmark, its scoring code or the model weights are public; the post links a demo but no download for this model.
• Per-scenario scores, so the null-handling case cannot be separated from the others.
• Throughput per GPU, latency, and the utilisation needed to reach the cost figures.
• Any test on documents from outside the training mix, such as your own contract or claim forms.

What a team can take from it

The pattern is a bounded task, a small model trained on it and a benchmark with fields that do not exist. That last case matters in production, where a model that invents a total is worse than one that returns nothing. If you run high-volume invoice, ID or form extraction through a frontier API, build that test from your own documents before any vendor does it for you.