olmOCR-Bench

Name: olmOCR-Bench Benchmark Results
Creator: Allen Institute for AI
License: https://creativecommons.org/licenses/by/4.0/

Allen Institute for AI

7,010 unit tests across 1,402 PDF documents. Tests parsing of tables, math, multi-column layouts, old scans, and more.

Models17

Papers28

Metrics9

SOTA History

Not enough data to show trend.

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 Base clean document parsing. Near-perfect	Editorial	99.9	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	olmocr-v0.3.0 #1 on headers/footers extraction	Editorial	95.1	2025	Source
2	chandra-ocr-0.1.0 Header/footer extraction	Editorial	90.8	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 Long documents with tiny text. #1 in category	Editorial	92.3	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	dots-ocr-3b #1 on table recognition	Editorial	88.3	2025	Source
2	chandra-ocr-0.1.0 Table recognition category. Near-best (dots.ocr: 88.3)	Editorial	88	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	marker-1.10.0 #1 on ArXiv paper parsing	Editorial	83.8	2025	Source
2	chandra-ocr-0.1.0 ArXiv paper parsing. Marker leads (83.8)	Editorial	82.2	2025	Source

Percentage of unit tests passed

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 7,010 unit tests across 1,402 PDF documents. #1 overall on olmOCR-Bench.	Editorial	83.1	2025	Source
2	infinity-parser-7b	Editorial	82.5	2025	Source
3	olmocr-v0.4.0	Editorial	82.4	2025	Source
4	paddleocr-vl	Editorial	80	2025	Source
5	Qianfan-OCR Baidu Qianfan-OCR 4B (Qwen3-4B + Qianfan-ViT), Apache 2.0, 192 langs. Layout-as-Thought.	Editorial	79.8	2026	Source
6	dots-ocr-3b	Editorial	79.1	2025	Source
7	mistral-ocr-3 Estimated based on 74% win rate vs OCR 2	Editorial	78	2025	Source
8	marker-1.10.0	Editorial	76.5	2025	Source
9	marker-1.10.1	Editorial	76.1	2025	Source
10	deepseek-ocr	Editorial	75.7	2025	Source
11	mineru-2.5	Editorial	75.2	2025	Source
12	mistral-ocr-api	Editorial	72	2025	Source
13	gpt-4o-anchored GPT-4o with anchored prompting	Editorial	69.9	2025	Source
14	nanonets-ocr2-3b	Editorial	69.5	2025	Source
15	gemini-flash-2	Editorial	63.8	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 Multi-column document parsing	Editorial	81.2	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 Mathematical notation in old scans. #1, leads by 5.4 points	Editorial	80.3	2025	Source
2	olmocr-v0.3.0 #2 on math in old scans	Editorial	79.9	2025	Source

Higher is better

Rank	Model	Source	Score	Year	Paper
1	chandra-ocr-0.1.0 Old scan recognition. #1 (GPT-4o: 40.7)	Editorial	50.4	2025	Source
2	gpt-4o #2 on old scans. Chandra leads by 9.7 points	Editorial	40.7	2025	Source