superPDF Performance
Free, no account, no upload: everything here runs in your browser, on your device. A 1 GB, 2,000-page PDF opens and edits in seconds, and superPDF’s own reader scores 53.7 on ParseBench across all 2,078 documents, instantly and with nothing to download.
Two ways superPDF reads a PDF
| Compared | superPDF reader (default) | High-accuracy reading for scans (optional) |
|---|---|---|
| Used for | Every PDF, automatically | Scanned pages, when you choose it |
| Download | None | 871 MB, once |
| Speed | Instant for digital PDFs; scans about 3 to 4 s a page | About 23 s a page |
| Best at | Digital PDFs: exact text, tables and charts drawn in the PDF | Scans, other scripts (Russian, Korean, Hebrew), charts that are images, hard tables |
| Words found on our 20-page scan | 67.9% | 81.7% |
| Tables (ParseBench) | 59.4 (all documents) | 75.4 (78-PDF sample) |
You don’t pick a reader: superPDF uses its own reader for every PDF, and offers high-accuracy reading only when a page is scanned. superPDF OCR is built on Tesseract, the open-source OCR engine.
Big files
1 GB, 2,000-page PDF (Chrome, Apple M4 Pro, 48 GB memory, 2026-10-11)
| Tool | Time | Peak memory |
|---|---|---|
| Redact PDF, 2 pages | About 10 s | 1.9 GB |
| Annotate PDF, 2 pages | About 10 s | 1.9 GB |
| Watermark PDF, chosen pages | About 8 s | 1.8 GB |
| Crop & resize, one page | About 9 s | 1.9 GB |
| Compare PDFs, two 2,000-page versions | About 72 s | 2.8 GB |
| PDF to text, Markdown, 2,000 text pages | 4.4 s | – |
Scanned pages
| Reader | Speed | Where |
|---|---|---|
| superPDF OCR (English) | 20 pages in 81 s, 50 pages in 142 s (3 to 4 s a page) | Every device, automatically |
| High-accuracy reading (LightOnOCR-3 0.8B) | About 23 s a page (11 to 50 s) | Computers, when you choose it; 871 MB download in about 25 s |
The page stays usable while it reads.
Words found on a 20-page mixed-language scan
| Reader | Words found |
|---|---|
| superPDF OCR | 67.9% |
| High-accuracy reading | 81.7% |
High-accuracy reading reads English, other European languages, Russian, Hebrew and Korean.
Accuracy
ParseBench (out of 100)
| Reader | Measured on | Overall | Tables | Charts | Content | Formatting | Grounding |
|---|---|---|---|---|---|---|---|
| superPDF reader (default) | All 2,078 documents | 53.7 | 59.4 | 31.7 | 76.3 | 47.9 | 53.3 |
| High-accuracy reading for scans (optional) | 78-PDF sample | 59.7 | 75.4 | 50.5 | 79.8 | 44.1 | 48.5 |
Both readers turn charts drawn in the PDF into data tables; high-accuracy reading also reads charts that are pictures.
Where superPDF stands
The superPDF reader next to leading parsers on ParseBench, out of 100. It reads charts drawn in a PDF, where most parsers score near zero.
| Parser | Kind | Overall | Tables | Charts | Content | Formatting | Grounding |
|---|---|---|---|---|---|---|---|
| LlamaParse Agentic Plus | Paid cloud service | 90.2 | 93.4 | 94.2 | 92.3 | 87.1 | 84.1 |
| Claude Opus 5.5 | Paid AI model | 78.0 | 93.9 | 64.1 | 91.8 | 77.0 | 63.2 |
| Gemini 3 Flash (thinking high) | Paid AI model | 75.0 | 91.5 | 64.8 | 90.9 | 68.3 | 59.8 |
| Mistral OCR 4 | Paid cloud service | 60.7 | 73.9 | 2.4 | 89.5 | 66.4 | 71.2 |
| Azure Document Intelligence | Paid cloud service | 59.6 | 86.0 | 1.6 | 84.9 | 51.9 | 73.8 |
| superPDF reader | Free, in your browser | 53.7 | 59.4 | 31.7 | 76.3 | 47.9 | 53.3 |
| PyMuPDF4LLM | Open source | 53.5 | 72.0 | 2.2 | 79.3 | 52.0 | 61.9 |
| Docling | Open source | 50.6 | 66.4 | 52.8 | 66.9 | 1.0 | 66.1 |
| Google Cloud Document AI | Paid cloud service | 50.4 | 55.1 | 1.4 | 83.7 | 50.5 | 61.3 |
| AWS Textract | Paid cloud service | 47.9 | 84.6 | 6.0 | 74.8 | 3.7 | 70.4 |
| MarkItDown | Open source | 18.6 | 15.8 | 2.0 | 64.5 | 0.9 | 9.9 |
Other parsers: the public ParseBench leaderboard, October 12, 2026. superPDF: our run of ParseBench’s open-source test suite on all 2,078 documents.
olmOCR-bench (42-PDF sample, 333 tests)
| Reader | Tests passed |
|---|---|
| High-accuracy reading for scans (optional) | 77.2% |
How we measured
- Machine: MacBook Pro, Apple M4 Pro, 48 GB of memory.
- Browser and dates: Chrome, October 11 and 12, 2026.
- Files: a 1 GB PDF of 2,000 scanned-style pages; a 2,000-page text-only report; a 20-page mixed-language scan.
- Benchmarks: ParseBench (LlamaIndex): the superPDF reader on all 2,078 documents with ParseBench’s open-source test suite (our run, October 12, 2026); high-accuracy reading on a 78-PDF sample (20 per dimension, seed 7). olmOCR-bench (Ai2): 42 PDFs, 333 tests.
- Leaderboard and LightOn’s scores: the leaderboard ranking (October 2026) and LightOn’s published 74.6 (ParseBench) and 85.5 (olmOCR-bench) are for LightOnOCR-3 0.8B in LightOn’s own full-precision setup, not superPDF’s scores. superPDF runs it in the browser with 4-bit weights and 1,540-pixel page images, converted to superPDF’s Markdown and tables: 59.7 and 77.2% on our samples.
- Languages: in our tests, high-accuracy reading did not read Gujarati well.
- Other devices: phones and 8 GB laptops are slower.