MULTILINGUAL · INTERNAL · TOP 43 LANGUAGES
Multilingual — 80.4% on Internal · top 43 languages.
There is no good public multilingual OCR benchmark, so we built one. It tests tables, math, ordering, layout, and text accuracy across the top 43 world languages — intentionally hard, to leave headroom. We also publish results on a 90-language long tail.
Ranked scoreboard.
Models ranked highest-to-lowest. Datalab variants in accent; competitors and prior generations in muted ink.
Rank Model Score
01 Datalab API 80.4%
02 Chandra 2 OSS 77.8%
03 Chandra 1 prior generation 69.4%
04 Gemini 2.5 Flash 67.6%
05 GPT-5 Mini 60.5%
+12.8 vs Gemini 2.5 Flash
We built this benchmark — no good public equivalent exists. Pairwise Bradley-Terry with Gemini-as-judge, randomized to reduce position bias, converted to ELO. 95% confidence intervals per row.
OTHER BENCHMARKS
All benchmarks →Compare on another doc type.
Tables Financial filings · regulatory PDFs · research data 90.7% Top of the public olmOCR-bench leaderboard. +2.7 vs Chandra 1. Handles colspan, rowspan, nested headers, merged cells. ArXiv Research papers · AI training corpora 90.4% 90.4% on ArXiv, +8.2 vs Chandra 1. Multi-column layout, inline equations, citation graphs — the substrate of most modern AI training data.
START
Gemini gets 67.6%. We get 80.4%.
Run a non-English PDF on Chandra — free tier, no credit card.