DeepSeek-OCR Is Strongest on Tables; PaddleOCR Only Recognized the Title on a Tax Form

Most OCR benchmarks give you only one overall score. It tells you who is ahead, but not exactly where each parser fails.

Akshay recently ran DeepSeek-OCR, PaddleOCR-VL, and Datalab through OmniParse Bench. This benchmark works differently from typical leaderboards: it breaks each page into a set of yes/no questions—whether this line of text is present, whether this cell is under the corresponding header, whether A comes before B. Each question also carries tags marking what makes the page difficult, such as handwriting, rotation, small font, forms, and tables.

His results on a 150-page subset:

• Datalab leads with an overall score of 88.2, mainly from text and layout
• DeepSeek-OCR is highest on tables, at 87.2
• On handwritten notes, Datalab passes 5/6 tests; the other two both score 0
• On a filled-out U.S. 1040 tax form, PaddleOCR-VL returned only the form title

The video shows page by page the original text, the outputs from the three parsers, and why each test passed or failed.

No one handles every document well. DeepSeek is strong on tables, but its text scores hold it back; PaddleOCR is stable on common pages, but it loses content when it encounters forms; Datalab's high overall score does not mean it leads in every subcategory. Before choosing a tool, first figure out what your documents mainly look like, then look at the scores under the corresponding tags.

The full version is much larger: 16,288 tests, 2,937 pages, 92 languages, open-source scoring code, and no LLM as a judge. It offers not just a leaderboard, but also a scorer and dataset you can use to test your own documents.

There is one detail that needs to be made clear. OmniParse Bench is published by Datalab, and in the full results Datalab's own model also ranks first, with Datalab accurate scoring 94.1. The code is open source, anyone can reproduce it, and the data itself is fine—but the publisher is also a participant, so keep that premise in mind when looking at the leaderboard.

For people working on document parsing, the real value of this kind of tool is not telling you who is the strongest. It lets you see where a given parser will break on your pages, rather than leaving you with only an overall score of unclear provenance.

GitHub repository: https://github.com/datalab-to/omni_parse_bench