DOCUMENT AI MODELS

Models built for documents, measured in public

Explore Nutrient models for verifying claims, detecting form fields, and classifying document images. Each model links to its public scorecard, demo, leaderboard, and evaluation data where available.

#1
of 16 systems on the public form-field benchmark (Typed-field F1 @ IoU 0.5)
.923
number-grounding ROC-AUC for grounding-en — vs. .48–.66 for general-purpose NLI baselines
4
specialized document models, each with a public demo, leaderboard, and benchmark dataset

Apache-2.0 open weights for grounding-en

Download it and run it yourself.

WHY SPECIALIZED MODELS

Narrow models, built to be checked

description="Frontier LLMs are strong generalists, but on precise document tasks, a small model tuned for the job wins — and wins in a way you can verify. Each Nutrient model does one document job well and is shaped to drop into an agent pipeline: ground a claim before a human reviews it, detect every field before a form is filled, label a page before it’s routed. Every result on this page is backed by a public dataset, a live demo, and a leaderboard you can rerun."

Grounding and verification

Cross-encoders that score whether document evidence actually supports a claim.

grounding-en grounding-multilingual

.923

ROC-AUC · number-grounding

.48–.66

general NLI baselines

CAVEAT

On general prose, the best listed baseline still edges the multilingual model (.961 vs. .926) — grounding’s edge is numbers, dates, and facts.

Form understanding

Detects every interactive field on a page — bounding box, type, label, radio-group identity — and returns JSON.

form-field-vlm

#1

Nutrient Hybrid · public 16-system leaderboard

#2

Nutrient VLM · public 16-system leaderboard

NOTE

Ranks first and second on the public form-field benchmark — see the full leaderboard below.

Document classification

Open-vocabulary: classifies document and figure images against labels you supply at runtime.

doc-img-classification

.880

broad top-1

.712

specialized top-1

CAVEAT

A listed baseline is slightly higher on synonym robustness.

The public form-field leaderboard
Typed-F1 @ IoU 0.5
0.477 #1 OF 16
VLM-only
0.406 #2 OF 16
Box recall
0.790 VS 0.364
Typed-field F1 @ IoU 0.5 — 100 pages, 701 fields
Nutrient Hybrid 0.48
Nutrient VLM 0.41
GPT-5.6 Sol 0.36
Gemini 3.1 Pro Preview 0.31
Gemini 3.5 Flash 0.27
GPT-5.6 Terra 0.26
GPT-5.6 Luna 0.17
Gemini 3.1 Flash-Lite 0.09
Claude 4.8 Opus 0.05
Qwen3-VL 4B 0.03
Claude 4.5 Haiku 0.02
Qwen3.5 4B (base) 0.01
Gemma 4 0.01
Claude 5 Sonnet 0.01
Pixtral 12B 0.00
Gemma 3 0.00

Typed-field F1 @ IoU 0.5 on 100 real pages/701 fields. At a looser IoU 0.2, the best generalist nearly ties on typed-F1 — the specialist’s edge is precise localization (box recall 0.790 vs. 0.364). Full methodology and the live leaderboard on Hugging Face.

RUN IT YOURSELF

Open weights, three lines

grounding-en is Apache-2.0 — pull it from Hugging Face. The multilingual, form-field, and classification models are commercial and served through Nutrient DWS Data Extraction.

from sentence_transformers import CrossEncoder


model = CrossEncoder("nutrientdocs/grounding-en")
score = model.predict([(claim, source_passage)])  # 0–1: is the claim supported?

IN THE PRODUCT

The same models, in your pipeline

Data Extraction API

form-field and grounding models power structured extraction with per-field confidence.
AI Assistant

grounding verifies generated answers against the source document.
Human-in-the-loop review

low-confidence fields are flagged for a reviewer instead of passed through.

MODEL CATALOG

Open weights where possible, public evidence throughout

The English grounding model is open under Apache-2.0. The multilingual grounding, form-field, and image-classification weights are commercial, with public model cards and reproducible evaluation benchmarks.

grounding-en

An open Apache-2.0 cross-encoder for verifying whether English document evidence supports a number, date, or factual claim. On the held-out benchmark, number-grounding ROC-AUC was .923 vs. .478–.658 for the listed general-purpose NLI baselines.

MODEL CARD
LIVE DEMO
LEADERBOARD
BENCHMARK DATASET
grounding-multilingual

A commercial, non-downloadable cross-encoder for grounding claims across 15+ languages. Its published benchmark reports .999 number-grounding ROC-AUC; on general prose, the best listed baseline scores higher (.961 vs. .926).

MODEL CARD
LIVE DEMO
LEADERBOARD
BENCHMARK DATASET
form-field-vlm

A commercial, non-downloadable VLM that returns form field boxes, types, labels, and radio-group identities. On the public 100-page, 701-field benchmark, the hybrid system scores .477 Typed F1 at IoU 0.5; the VLM alone scores .406.

MODEL CARD
LIVE DEMO
LEADERBOARD
BENCHMARK DATASET
doc-img-classification

A commercial, non-downloadable open-vocabulary model that classifies document and figure images against labels supplied at runtime. On the public ~48-label benchmark, it reports .880 broad and .712 specialized top-1 accuracy; a listed baseline is slightly higher on synonym robustness.

MODEL CARD
LIVE DEMO
LEADERBOARD
BENCHMARK DATASET

Frequently asked questions

Are the models open source?

grounding-en is Apache-2.0 open weights; multilingual grounding, form-field-vlm, and doc-img-classification are commercial, with public model cards and reproducible benchmarks.

How were the benchmarks run?

On public datasets with published methodology; each model links to its dataset and a live leaderboard Space you can rerun.

How do I run them?

Open weights via Hugging Face (transformers/sentence-transformers); commercial models through Nutrient DWS Data Extraction.

How do these fit an agent workflow?

They verify and structure before a human step: ground a claim, detect every field, and classify a page — so an agent hands off checked output.

Which languages does grounding cover?

grounding-en is English; grounding-multilingual covers 15+ languages.

DOCUMENT AI Put benchmarked document models in your pipeline

Talk to Nutrient about document AI, or try the same grounding and form-field capabilities through DWS Data Extraction.

Try DWS Data Extraction