MinerU against plain pdftotext on Vietnam economic reports
MinerU turns PDFs into Markdown with layout models and OCR. I needed clean numbers out of Vietnam economic reports for my project, and on 14 October 2026 I checked whether MinerU 4 reads them better than pdftotext -layout, the plain tool from poppler. On PDFs that have a text layer it doesn't. It only wins on scans.
How I tested
- Three real sources: a World Bank report on Vietnam, the IMF's 2025 Article IV report on Vietnam, and the Vietnamese statistics office's Q3 2026 report, in Vietnamese.
- 60 numbers from tables, each checked against the page image for its value, row and column.
- 200 Vietnamese words from a text PDF.
- One scanned table (IMF Table 4, a bitmap with 8 numbers) and one page rendered to PNG so it had to be OCR'd (100 words).
- MinerU's flash and basic tiers against
pdftotext -layout, all run locally on a Mac.
Results
14 October 2026, run locally on a Mac:
- Table numbers right, text PDFs (of 60): pdftotext 57; MinerU flash 43; MinerU basic 57
- Scanned table numbers right (of 8): pdftotext 0; MinerU flash 0; MinerU basic 6
- Vietnamese words right, text PDF (of 200): pdftotext 200; MinerU flash 200; MinerU basic 200
- Vietnamese words right, OCR of an image (of 100): pdftotext 0; MinerU flash 53; MinerU basic 53
- Seconds per page: pdftotext 0.002; MinerU flash 0.10 to 0.18; MinerU basic 0.69 to 1.43
On text PDFs, MinerU's best local tier tied pdftotext at several hundred times the time per page. Its faster tier got 14 more table numbers wrong. On the scan MinerU was the only one that read anything, but 6 of 8 numbers is 75%, and I wanted 95% before trusting a number. From an image it read 53% of the Vietnamese words.
What I do now
I keep pdftotext for PDFs with a text layer. When the statistics office also publishes .docx or .xlsx files, I read those directly instead of going through the PDF.
What I did not test
MinerU's standard tier, and old government PDFs that use legacy Vietnamese fonts (TCVN3, VNI). Those are the cases where I would test it again. The sample is small: one wrong number moves a score by 1.7 points.