Codesota · OCR · Benchmarks · olmOCR-BenchHome/OCR/Benchmarks/olmOCR-Bench
Allen Institute for AI

olmOCR-Bench.

PDF content extraction benchmark with 7,010 unit tests across 1,402 PDF documents.

View on AlphaXiv
§ 01 · base

base.

Higher is better

#ModelScoreSource
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
99.9codesota-api
2
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
99.7codesota-api
3
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
99.6codesota-api
4
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
99.6codesota-api
§ 02 · headers-footers

headers-footers.

Higher is better

#ModelScoreSource
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
96.1codesota-api
2
olmocr-v0.3.0
Fetched from CodeSOTA API on 2026-04-20
95.1codesota-api
3
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
90.8codesota-api
4
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
42codesota-api
§ 03 · long-tiny-text

long-tiny-text.

Higher is better

#ModelScoreSource
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
92.3codesota-api
2
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
91.4codesota-api
3
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
81.9codesota-api
4
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
80.4codesota-api
§ 04 · multi-column

multi-column.

Higher is better

#ModelScoreSource
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
92.2codesota-api
2
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
84.8codesota-api
3
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
83.7codesota-api
4
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
81.2codesota-api
§ 05 · arxiv

arxiv.

Higher is better

#ModelScoreSource
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
89.6codesota-api
2
marker-1.10.0
Fetched from CodeSOTA API on 2026-04-20
83.8codesota-api
3
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
83codesota-api
4
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
82.2codesota-api
5
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
80.1codesota-api
§ 06 · tables

tables.

Higher is better

#ModelScoreSource
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
89codesota-api
2
dots-ocr-3b
Fetched from CodeSOTA API on 2026-04-20
88.3codesota-api
3
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
88codesota-api
4
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
84.9codesota-api
5
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
81.6codesota-api
§ 07 · old-scans-math

old-scans-math.

Higher is better

#ModelScoreSource
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
85.6codesota-api
2
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
82.3codesota-api
3
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
80.3codesota-api
4
olmocr-v0.3.0
Fetched from CodeSOTA API on 2026-04-20
79.9codesota-api
§ 08 · Pass Rate

Pass Rate.

Percentage of unit tests passed

Higher is better

#ModelScoreSource
infinity-parser2-pro
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for infly/Infinity-Parser2-Pro. Source: Infinity Parser technical report / HF leaderboard entry; metric: Accuracy.; PWC evaluation id 4945; paper: Infinity-Parser2-Pro
87.6%paperswithcode-public-api
2
chandra-2
Mapped from PWC olmOCR-Bench Accuracy.; Chandra OCR 2 (5B params, Qwen3.5 backbone) reported on olmOCR-Bench in the HF model card; Overall score 85.9 +/- 0.8. Sub-categories on the same benchmark: ArXiv 90.2, Old Scans Math 89.3, Tables 89.9, Old Scans 49.8, Headers and Footers 92.5, Multi column 83.5, Long tiny text 92.1, Base 99.6. Source: own benchmarks, reported in the chandra-ocr-2 HF model card.; PWC evaluation id 1165; paper: Chandra OCR 2
85.9%paperswithcode-public-api
3
dots.mocr
Mapped from PWC olmOCR-Bench Accuracy.; Overall score on olmOCR-Bench from the dots.mocr Hugging Face model card (section 1.2). 3B image-text-to-text VLM; page-header and page-footer cells deleted from the result markdown per the card's note; metric source: olmocr plus internal evaluations.; PWC evaluation id 1148; paper: Multimodal OCR: Parse Anything from Documents
83.9%paperswithcode-public-api
4
LightOnOCR-2-1B
Mapped from PWC olmOCR-Bench Accuracy.; Overall score on olmOCR-Bench reported on the Hugging Face model card. Score excludes the Headers/Footers (H&F) sub-test because that category rewards omission rather than transcription, while LightOnOCR-2 is trained for full-page transcription and intentionally preserves headers/footers.; PWC evaluation id 1147; paper: LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR
83.2%paperswithcode-public-api
5
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
83.1%codesota-api
6
chandra
Mapped from PWC olmOCR-Bench Accuracy.; Reported in the datalab-to/chandra Hugging Face model card on olmOCR-Bench. Overall score 83.1 +/- 0.9; sub-category scores from the same table: ArXiv 82.2, Old Scans Math 80.3, Tables 88.0, Old Scans 50.4, Headers and Footers 90.8, Multi column 81.2, Long tiny text 92.3, Base 99.9. Source: own benchmarks.; PWC evaluation id 1327; paper: Chandra
83.1%paperswithcode-public-api
7
infinity-parser-7b
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for infly/Infinity-Parser-7B. Source: Infinity Parser technical report / HF leaderboard entry; metric: Accuracy.; PWC evaluation id 4946; paper: Infinity Parser: Layout Aware Reinforcement Learning for Scanned Document Parsing
82.5%paperswithcode-public-api
8
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
82.4%codesota-api
9
olmocr-2-7b-1025-7b
Mapped from PWC olmOCR-Bench Accuracy.; PWC evaluation id 679; paper: olmOCR 2: Unit Test Rewards for Document OCR
82.4%paperswithcode-public-api
10
falcon-ocr
Mapped from PWC olmOCR-Bench Accuracy.; olmOCR-Bench accuracy reported by tiiuae/Falcon-OCR model card (arXiv:2603.27365, Apr 2026). 'Average' across 8 category splits: ArXiv Math 80.5, Base 99.5, Headers/Footers 94.0, Long Tiny Text 78.5, Multi Column 87.1, Old Scans 43.5, Old Scans Math 69.2, Tables 90.3. Inference via Layout + OCR two-stage pipeline (PP-DocLayoutV3 layout detection + Falcon-OCR category-prompted VLM).; PWC evaluation id 1156; paper: Falcon Perception
80.3%paperswithcode-public-api
11
paddleocr-vl
Mapped from PWC olmOCR-Bench Accuracy.; PWC evaluation id 86; paper: PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model
80%paperswithcode-public-api
12
Qianfan-OCR
Mapped from PWC olmOCR-Bench Accuracy.; End-to-end OCR on olmOCR-Bench; overall 79.8 (Arxiv Math 80.1, Old Scans Math 73.1, Table Tests 81.6, Old Scans 42.0, Multi Column 80.4, Long Tiny Text 89.1, Headers Footers 92.2).; PWC evaluation id 1196; paper: Qianfan-OCR: A Unified End-to-End Model for Document Intelligence
79.8%paperswithcode-public-api
13
Qwen3-VL-4B
Fetched from CodeSOTA API on 2026-04-20
79.2%codesota-api
14
PaddleOCR-VL-1.5
Fetched from CodeSOTA API on 2026-04-20
79.1%codesota-api
15
dots-ocr-3b
Fetched from CodeSOTA API on 2026-04-20
79.1%codesota-api
16
dots-ocr
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for rednote-hilab/dots.ocr. Source: dots.ocr technical report / HF leaderboard entry; metric: Accuracy.; PWC evaluation id 4947; paper: dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model
79.1%paperswithcode-public-api
17
mistral-ocr-3
Fetched from CodeSOTA API on 2026-04-20
78%codesota-api
18
mineru-2.5
Mapped from PWC olmOCR-Bench Accuracy.; PWC evaluation id 93; paper: MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing
77.5%paperswithcode-public-api
19
marker-1.10.0
Fetched from CodeSOTA API on 2026-04-20
76.5%codesota-api
20
deepseek-ocr-2
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for deepseek-ai/DeepSeek-OCR-2. Source: HF leaderboard entry and DeepSeek-OCR-2 arXiv-linked model card; metric: Accuracy.; PWC evaluation id 4948; paper: DeepSeek-OCR 2: Visual Causal Flow
76.3%paperswithcode-public-api
21
marker-1.10.1
Fetched from CodeSOTA API on 2026-04-20
76.1%codesota-api
22
lightonocr-1b-1025
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for lightonai/LightOnOCR-1B-1025. Source note: Headers & Footers category excluded; metric: Accuracy.; PWC evaluation id 4949; paper: LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR
76.1%paperswithcode-public-api
23
MonkeyOCR-pro-3B
Fetched from CodeSOTA API on 2026-04-20
75.8%codesota-api
24
deepseek-ocr
Fetched from CodeSOTA API on 2026-04-20
75.7%codesota-api
25
DeepSeek-OCR
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for deepseek-ai/DeepSeek-OCR. Source: olmOCR-Bench GitHub leaderboard entry and DeepSeek-OCR arXiv-linked model card; metric: Accuracy.; PWC evaluation id 4950; paper: DeepSeek-OCR: Contexts Optical Compression
75.7%paperswithcode-public-api
26
olmocr
Mapped from PWC olmOCR-Bench Accuracy.; Paper Table 4, olmOCR-Bench overall unit-test pass rate for Ours (v0.1.75 Anchored). Category scores reported in the paper: AR 74.9, OSM 71.2, TA 71.0, OS 42.2, HF 94.5, MC 78.3, LTT 73.3, Base 98.3.; PWC evaluation id 4979; paper: olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language Models
75.5%paperswithcode-public-api
27
mineru-2.5
Fetched from CodeSOTA API on 2026-04-20
75.2%codesota-api
28
GLM-OCR
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for zai-org/GLM-OCR. Source note: Headers & Footers category excluded; evaluated via ZAI API; metric: Accuracy.; PWC evaluation id 4951; paper: GLM-OCR Technical Report
75.2%paperswithcode-public-api
29
mistral-ocr-api
Fetched from CodeSOTA API on 2026-04-20
72%codesota-api
30
firered-ocr
Mapped from PWC olmOCR-Bench Accuracy.; Overall score reported on the official Hugging Face olmOCR-bench leaderboard for FireRedTeam/FireRed-OCR. Source note: Headers & Footers category excluded; metric: Accuracy.; PWC evaluation id 4952; paper: FireRed-OCR Technical Report
70.2%paperswithcode-public-api
31
gpt-4o-anchored
Fetched from CodeSOTA API on 2026-04-20
69.9%codesota-api
32
nanonets-ocr2-3b
Fetched from CodeSOTA API on 2026-04-20
69.5%codesota-api
33
gemini-flash-2
Fetched from CodeSOTA API on 2026-04-20
63.8%codesota-api
§ 09 · old-scans

old-scans.

Higher is better

#ModelScoreSource
Qianfan-OCR
Fetched from CodeSOTA API on 2026-04-20
73.1codesota-api
2
chandra-ocr-0.1.0
Fetched from CodeSOTA API on 2026-04-20
50.4codesota-api
3
olmocr-v0.4.0
Fetched from CodeSOTA API on 2026-04-20
47.7codesota-api
4
LightOnOCR-2-1B
Fetched from CodeSOTA API on 2026-04-20
42.2codesota-api
5
gpt-4o
Fetched from CodeSOTA API on 2026-04-20
40.7codesota-api
§ Related · Explore

More OCR content.

Verified Model Reviews
Comparisons & Guides
View all OCR benchmarks → Back to All Benchmarks