Benchmark Results
EdgeParse publishes two scoreboards. They use different overall formulas — do not mix numbers across boards.
| Board | Export | Formula | Role |
|---|---|---|---|
| EdgeParse harness | benchmarkSnapshot | Product overall (quality + structure) | Regression / product score |
| Official OpenDataLoader board (odl-bench) | odlBenchSnapshot | mean(NID, TEDS, MHS) | Match public OpenDataLoader leaderboard |
Board A — EdgeParse harness
Section titled “Board A — EdgeParse harness”Updated 2026-10-02 · Apple M4 Max · 200 docs · overall 0.760 · 0.284 s/doc
| Tool | NID | TEDS | MHS | Overall | Speed |
|---|---|---|---|---|---|
| EdgeParse | 0.880 | 0.562 | 0.478 | 0.760 | 0.284 s/doc |
| EdgeParse [hybrid] | 0.863 | 0.603 | 0.506 | 0.767 | 0.810 s/doc |
| Docling (IBM) | 0.875 | 0.568 | 0.450 | 0.758 | 2.332 s/doc |
| OpenDataLoader | 0.873 | 0.320 | 0.441 | 0.733 | 0.022 s/doc |
| PyMuPDF4LLM | 0.860 | 0.509 | 0.411 | 0.732 | 0.640 s/doc |
| OpenDataLoader [hybrid] | 0.869 | 0.422 | 0.411 | 0.731 | 2.537 s/doc |
| MarkItDown | 0.807 | 0.193 | 0.001 | 0.564 | 0.189 s/doc |
Harness takeaways
Section titled “Harness takeaways”- Deterministic EdgeParse leads this board at overall 0.760 and 0.284 s/doc
- 8× faster than Docling · 2× faster than PyMuPDF4LLM · 0.1× faster than OpenDataLoader
- No ML stack required for born-digital PDFs; optional in-browser OCR for image tables
Board B — Official odl-bench
Section titled “Board B — Official odl-bench”Updated 2026-10-01 · Apple M4 Max · 200 docs · formula {claims.odl.formula}
EdgeParse [hybrid] leads NID, TEDS, MHS, and overall (0.900).
| Tool | NID | TEDS | MHS | Overall | Status |
|---|---|---|---|---|---|
| EdgeParse [hybrid] | 0.917 | 0.928 | 0.836 | 0.900 | finished |
| Docling | 0.905 | 0.925 | 0.827 | 0.891 | finished |
| OpenDataLoader [hybrid] | 0.915 | 0.676 | 0.822 | 0.877 | finished |
| PyMuPDF4LLM | 0.916 | 0.803 | 0.772 | 0.873 | finished |
| EdgeParse | 0.894 | 0.773 | 0.763 | 0.857 | finished |
| EdgeParse [wasm] | 0.891 | 0.776 | 0.761 | 0.854 | finished |
| OpenDataLoader | 0.912 | 0.483 | 0.757 | 0.842 | finished |
| MarkItDown | 0.844 | 0.273 | 0.000 | 0.589 | finished |
| Unstructured | 0.093 | 0.000 | 0.054 | 0.086 | partial |
odl-bench takeaways
Section titled “odl-bench takeaways”- EdgeParse [hybrid] — overall 0.900 (leads every metric)
- EdgeParse (deterministic) — overall 0.857
- EdgeParse [wasm] — overall 0.854 (TEDS 0.776)
- Unstructured is partial (not a finished competitor); Marker / MinerU / Nutrient were not run
Speed Comparison (harness)
Section titled “Speed Comparison (harness)”| Comparison | Factor |
|---|---|
| EdgeParse vs Docling | 8.2× faster |
| EdgeParse vs PyMuPDF4LLM | 2.3× faster |
| EdgeParse vs OpenDataLoader | 0.1× faster |
| EdgeParse vs MarkItDown | 0.7× faster |
Test Environment
Section titled “Test Environment”- Hardware: Apple M4 Max
- Corpus: 200 diverse PDF documents
- Harness path:
benchmark/ - Official board path:
odl-bench/ - Categories: Academic papers, financial reports, invoices, government forms, books, manuals
See It in Action
Section titled “See It in Action”- Try Live Demo — parse any PDF in your browser with WebAssembly, no install needed
- Enterprise — production-grade deployment with priority support by Elitizon