Saturating the olmOCR benchmark
What it means when a benchmark stops being hard — and where we look next.
THE DOCEN TEAM·Jan 28, 2026·6 MIN READ
What it means when a benchmark stops being hard — and where we look next.
What we measure
We publish benchmarks because a document intelligence model is only as good as its behavior on documents you can't cherry-pick. We test recognition, tables, layout, and extraction against sensible baselines.
- Character and word error rate on scanned and photographed pages.
- Table-cell accuracy on complex, merged layouts.
- Field-level F1 for schema extraction, with citations checked.
- Latency per page at production settings.
Results
0.42%
character error rate
98.7%
table-cell accuracy
94.3%
extraction F1
Method
Every number here comes from documents held out of training. We report the settings, keep the evaluation reproducible, and update the figures as models change. When a result looks too good, we assume the test is wrong until we've checked it.
Want to see it on your own documents? Open the playground or reach out — we're happy to run a sample with you.
benchmarksOCRevaluation