Name: Intelligent Document Processing Leaderboard
Creator: Nanonets
License: https://opensource.org/licenses/MIT

Question 1

What is the best document AI model in 2026?

Accepted Answer

As of March 2026, Gemini 3.1 Pro by Google leads the IDP Leaderboard with an overall score of 83.2%, followed by Nanonets OCR2+ (81.8%) and Gemini-3-Pro (81.4%). These scores are averaged across three benchmarks: OlmOCR (OCR quality), OmniDocBench (document parsing), and IDP Core (extraction + VQA).

Question 2

What benchmarks are used in the IDP Leaderboard?

Accepted Answer

The IDP Leaderboard evaluates models on three complementary benchmarks: (1) OlmOCR Bench — tests OCR fidelity on math equations, tables, headers/footers, and reading order across ~7,000 document pages; (2) OmniDocBench v1.5 — comprehensive document parsing evaluation covering text extraction, formula recognition (CDM), table structure (TEDS), and reading order; (3) IDP Core Bench — real-world document processing tasks including key information extraction (KIE), OCR on handwriting, table understanding, and visual question answering (VQA).

Question 3

How are document AI models scored on the IDP Leaderboard?

Accepted Answer

Each model is scored on three benchmarks independently. The overall score is the unweighted mean of the three benchmark scores. Within each benchmark, task-specific metrics are used: OlmOCR uses pass/fail tests, OmniDocBench uses edit distance and TEDS, and IDP Core uses exact-match and similarity metrics. All models are evaluated with identical prompts, images, and scoring pipelines for fair comparison.

Question 4

What is intelligent document processing (IDP)?

Accepted Answer

Intelligent Document Processing (IDP) is the use of AI to automatically read, understand, and extract information from documents such as invoices, receipts, contracts, and forms. IDP combines OCR (optical character recognition), table extraction, key information extraction, and visual question answering to convert unstructured documents into structured, actionable data.

Question 5

Which document AI model is best for OCR?

Accepted Answer

For OCR tasks, Nanonets OCR2+ by Nanonets achieves the highest score of 82.2% on the OlmOCR benchmark, which evaluates math equation rendering, table structure preservation, header/footer handling, and multi-column reading order.

Question 6

Which document AI model is best for table extraction?

Accepted Answer

For table extraction, Gemini-3-Flash by Google achieves the highest TEDS score of 87.7% on OmniDocBench, measuring table structure and content accuracy.

Question 7

Is the IDP Leaderboard open source?

Accepted Answer

Yes. The evaluation code, benchmark datasets, and scoring pipelines are open source and available on GitHub at https://github.com/NanoNets/docext. All results are reproducible. The leaderboard is sponsored by Nanonets.

#	Model	Overall	OlmOCR	OmniDoc	IDP
1	Gemini 3.1 ProGoogle	83.2	74.6	85.3	89.6
2	Nanonets OCR2+Nanonets	81.8	82.2	89.5	73.8
3	Gemini-3-ProGoogle	81.4	73.5	88.8	81.8
4	GPT-5.4OpenAI	81.0	73.4	85.3	84.4
5	Claude Sonnet 4.6Anthropic	80.8	74.4	86.9	81.2
6	Claude Opus 4.6Anthropic	80.3	73.9	85.9	81.1
7	Gemini-3-FlashGoogle	79.9	69.2	90.1	80.5
8	GPT-5.2OpenAI	79.2	72.2	88.0	77.4
9	Qwen3.5-9BAlibaba	77.0	78.1	76.7	76.2
10	Qwen3.5-4BAlibaba	73.1	77.2	67.6	74.5
11	GPT-5-MiniOpenAI	70.8	56.7	82.5	73.3
12	GPT-4.1OpenAI	70.0	55.5	79.9	74.7
13	Claude Haiku 4.5Anthropic	69.6	56.2	79.6	72.9
14	Ministral-8BMistral AI	69.3	57.8	78.3	71.7
15	GLM-OCRZhipu AI	63.6	66.7	69.2	54.9
16	Qwen3.5-2BAlibaba	63.2	73.7	48.7	67.1
17	Qwen3.5-0.8BAlibaba	58.0	65.6	47.3	61.2
18	GPT-5-NanoOpenAI	50.7	22.8	63.4	65.8
19	Llama-3.2-Vision-11BMeta	50.1	47.2	44.6	58.6
20	Pixtral-12BMistral AI	46.0	36.8	42.3	59.0

Intelligent Document Processing Leaderboard

About the Leaderboard