DocParsingBench - Text: leaderboard

Systems turn 1,400 pages of real business documents (finance, legal, research, manufacturing, education; Chinese and English) into Markdown; the README does not state how many pages each system was scored on. This board is the text part: half one minus normalized edit distance on text lines, half CDM on inline formulas, after Hungarian matching of segments.

Metric: Text score: 0.5 (1 - NED) + 0.5 inline formula CDM (0-1). Source: github.com. Saturation forecast: Rough model projection: around 2026. 17 models tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1OXR-1.00.906
2Jina-OCR-v10.8905
3PaddleOCR-VL-1.60.8886
4FireRed-OCR0.8802
5Dots.mocr0.8673
6Infinity-Parser2-Pro0.8643
7TeleOCR0.8642
8Chandra-OCR-20.8624
9MinerU2.5-Pro0.8622
10OvisOCR20.8576
11Qianfan-OCR0.853
12ABot-OCR0.8498
13GLM-OCR0.8372
14Youtu-Parsing0.8134
15HunyuanOCR 1.50.7873

Interactive version: theaggregate.ai/benchmark?slug=docparsingbench-text · How It Works · Data refreshed daily, snapshot 2026-11-16.