MarathonDocBench
Models parse 100 documents spanning 103 to 300 pages. Evaluation scores content, cross-page continuations, figures, reading order, and heading hierarchy.
The paper describes about 205,000 layout elements. The linked Hugging Face dataset returned HTTP 401 during this audit, so public data access remains unverified.
- Made by
- Baidu ERNIE Team
- Released
- October 2026
- Kind
- Documents, charts and text in images
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Short answer
- Measures
- Capability
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- One fixed set
- Run by
- A host
- In views
- Image, video and 3D generation
- Strongest model tested
-
Gemini 3.8 Flash #24 on the capability index
- Last in a launch post
- October 2026
- Last model added
- October 2026
- Scores refreshed
- November 20th, 2026
Unlimited OCR 2 leads with 77.1%
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best.
Unlimited OCR 2 leads with 77.1%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model. Striped bars are claims, and dotted bars are reruns by Epoch AI or another board.
- 1
Unlimited OCR 2 77.1%
- 2 Kimi K3 33.7%
- 3
Gemini 3.8 Flash 28.6%
- 4
Qwen3.8-Flash-Next 27.8%
- 5
Gemini 3.1 Pro 27%
- 6
Qwen 3.8 Max 22.9%
- 7
MiniMax-M3 18%
- 8
DeepSeek V4.1 Flash 12.3%
- 9 GLM-5.3-Flash 7.51%
Source: The MarathonDocBench paper
- 1
Unlimited OCR 2 77.1%
- 2 Kimi K3 33.7%
- 3
Gemini 3.8 Flash 28.6%
- 4
Qwen3.8-Flash-Next 27.8%
- 5
Gemini 3.1 Pro 27%
- 6
Qwen 3.8 Max 22.9%
- 7
MiniMax-M3 18%
- 8
DeepSeek V4.1 Flash 12.3%
- 9 GLM-5.3-Flash 7.51%
Sources: Labs’ model cards and the MarathonDocBench paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 77.1% | Oct 2026 | Official leaderboard | |
| 2 | Kimi K3 | 33.7% | Jul 2026 | Official leaderboard | |
| 3 | | 28.6% | Sep 2026 | Official leaderboard | |
| 4 | | 27.8% | Aug 2026 | Official leaderboard | |
| 5 | | 27% | Feb 2026 | Official leaderboard | |
| 6 | | 22.9% | Aug 2026 | Official leaderboard | |
| 7 | | 18% | Jun 2026 | Official leaderboard | |
| 8 | | 12.3% | Sep 2026 | Official leaderboard | |
| 9 | GLM-5.3-Flash | 7.51% | Aug 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | 77.1% | Oct 2026 | Lab model card | ||
| 2 | Kimi K3 | 33.7% | Jul 2026 | Official leaderboard | |
| 3 | 28.6% | Sep 2026 | Official leaderboard | ||
| 4 | 27.8% | Aug 2026 | Official leaderboard | ||
| 5 | 27% | Feb 2026 | Official leaderboard | ||
| 6 | 22.9% | Aug 2026 | Official leaderboard | ||
| 7 | 18% | Jun 2026 | Official leaderboard | ||
| 8 | 12.3% | Sep 2026 | Official leaderboard | ||
| 9 | GLM-5.3-Flash | 7.51% | Aug 2026 | Official leaderboard |
Labs report scores for 1 model in their cards. Switch the scores above to rank them together.
Rankings by part
We read 8 parts from the benchmark’s paper, each ranking models on one part of the test, such as Text (Edit), T-Text (Edit) and Formula (CDM). Unlimited OCR 2 leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.
Text (Edit)
Normalized edit distance from 0 to 1 9 models we track have a score here.
- 1
Unlimited OCR 2 0.10
- 2 Kimi K3 0.47
- 3
Gemini 3.1 Pro ↑2 0.52
- 4
Gemini 3.8 Flash ↓1 0.52
- 5
Qwen3.8-Flash-Next ↓1 0.54
- 6
Qwen 3.8 Max 0.60
- 7
MiniMax-M3 0.66
- 8
DeepSeek V4.1 Flash 0.74
- 9 GLM-5.3-Flash 0.86
T-Text (Edit)
Normalized edit distance from 0 to 1 9 models we track have a score here.
- 1
Unlimited OCR 2 0.12
- 2
Gemini 3.8 Flash ↑1 0.66
- 3 Kimi K3 ↓1 0.66
- 4
Qwen3.8-Flash-Next 0.75
- 5
Gemini 3.1 Pro 0.76
- 6
Qwen 3.8 Max 0.80
- 7
MiniMax-M3 0.82
- 8
DeepSeek V4.1 Flash 0.93
- 9 GLM-5.3-Flash 0.94
Formula (CDM)
9 models we track have a score here.
- 1
Unlimited OCR 2 87.9%
- 2 Kimi K3 40.5%
- 3
Gemini 3.1 Pro ↑2 30.4%
- 4
Qwen3.8-Flash-Next 28.3%
- 5
Gemini 3.8 Flash ↓2 26.4%
- 6
Qwen 3.8 Max 18.8%
- 7
MiniMax-M3 14.2%
- 8
DeepSeek V4.1 Flash 10.4%
- 9 GLM-5.3-Flash 4.50%
Table (TEDS)
9 models we track have a score here.
- 1
Unlimited OCR 2 74.8%
- 2
Gemini 3.1 Pro ↑3 28.5%
- 3 Kimi K3 ↓1 27.8%
- 4
Qwen3.8-Flash-Next 25.4%
- 5
Gemini 3.8 Flash ↓2 23.4%
- 6
MiniMax-M3 ↑1 17.8%
- 7
Qwen 3.8 Max ↓1 15.2%
- 8
DeepSeek V4.1 Flash 12.7%
- 9 GLM-5.3-Flash 7.12%
T-Table (TEDS)
9 models we track have a score here.
- 1
Unlimited OCR 2 71.1%
- 2 Kimi K3 26.3%
- 3
Gemini 3.1 Pro ↑2 21.7%
- 4
Qwen3.8-Flash-Next 21%
- 5
Gemini 3.8 Flash ↓2 19.6%
- 6
MiniMax-M3 ↑1 14.8%
- 7
Qwen 3.8 Max ↓1 13.7%
- 8
DeepSeek V4.1 Flash 12.6%
- 9 GLM-5.3-Flash 6.65%
Figure (F1)
9 models we track have a score here.
- 1
Unlimited OCR 2 71.8%
- 2
Qwen 3.8 Max ↑4 33.4%
- 3
Qwen3.8-Flash-Next ↑1 24.1%
- 4 Kimi K3 ↓2 21.4%
- 5
Gemini 3.1 Pro 18.9%
- 6
Gemini 3.8 Flash ↓3 18.6%
- 7 GLM-5.3-Flash ↑2 7.10%
- 8
DeepSeek V4.1 Flash 6.95%
- 9
MiniMax-M3 ↓2 3.67%
ReadOrder (Edit)
Normalized edit distance from 0 to 1 9 models we track have a score here.
- 1
Unlimited OCR 2 0.15
- 2 Kimi K3 0.61
- 3
Gemini 3.1 Pro ↑2 0.65
- 4
Qwen3.8-Flash-Next 0.68
- 5
Gemini 3.8 Flash ↓2 0.69
- 6
Qwen 3.8 Max 0.73
- 7
MiniMax-M3 0.76
- 8
DeepSeek V4.1 Flash 0.84
- 9 GLM-5.3-Flash 0.92
Heading (TEDS)
9 models we track have a score here.
- 1
Unlimited OCR 2 48.2%
- 2 Kimi K3 28.9%
- 3
Gemini 3.8 Flash 27.3%
- 4
Qwen3.8-Flash-Next 21.5%
- 5
MiniMax-M3 ↑2 17.2%
- 6
Qwen 3.8 Max 16.1%
- 7
Gemini 3.1 Pro ↓2 9.68%
- 8
DeepSeek V4.1 Flash 7.62%
- 9 GLM-5.3-Flash 6.55%
2 models have held number one in the benchmark’s paper since its release in October 2026 2 models have held number one with claims included since its release in October 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The MarathonDocBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official MarathonDocBench leaderboard and labs’ model cards
One lab put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Oct 27, 2026 | 77.1 | MarathonDocBench Overall |
One lab claim next to the leaderboard
Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. 1 of them match, which usually means one side copied the other. Every benchmark.
| Model | Lab says | Board says | Gap |
|---|---|---|---|
| | 77.1% MarathonDocBench Overall | 77.1% Leaderboard | Same |
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
10 of 11 checks pass, 1 warning. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark
Catalog record 1 warning 1 pass
- Warning Task access We don’t know whether the tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the benchmark’s paper (9 entries) and launch posts (1 score).
- Pass Models scored 9 models we track.
- Pass Newest score The newest is from October 2026, for Unlimited OCR 2.
- Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3.8 Flash (#24).
- Pass Headroom The best is 77.1%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 9 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 2 pass
- Pass Lab claims The one pair agrees within 10 points.
- Pass Known issues No published audit found problems.