MarathonDocBench: leaderboard
Each system parses 100 documents of 103 to 300 pages end to end into text, formulas, tables, figures, reading order and headings, including text and tables that continue across pages; the score is the paper's overall aggregate. MPDocBench, its nearest neighbour, uses documents of up to 30 pages.
Metric: Overall (0-100). Source: arxiv.org. Saturation forecast: Around May 2028. 9 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Unlimited-OCR-2 | 77.06 | |
| 2 | Kimi K3 | 33.74 | #56 |
| 3 | Gemini 3.8 Flash | 28.61 | #36 |
| 4 | Qwen3.8-Flash-Next | 27.83 | #117 |
| 5 | Gemini 3.1 Pro (Preview) | 26.97 | #93 |
| 6 | Qwen 3.8 Max | 22.93 | #59 |
| 7 | MiniMax-M3 | 17.97 | #272 |
| 8 | DeepSeek V4.1 Flash | 12.34 | #62 |
| 9 | GLM-5.3 Flash | 7.51 | #125 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=marathondocbench · How It Works · Data refreshed daily, snapshot 2026-11-16.