GlotOCR Bench: leaderboard

Metric: Acc@5 (%): share of sentences transcribed with CER at most 0.05, macro-averaged over scripts, mean of the high (Latin), mid (9 scripts) and low-resource (148 scripts) tiers, on GlotOCR Bench (16,375 real multilingual sentences rendered as clean images across 158 Unicode scripts; zero-shot transcription; CER is the character edit distance with whitespace ignored, capped at 1, best of four output normalizations); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 14 models tracked.

Top models

#ModelScore
1Gemini 3.1 Flash Lite61.9
2GPT-4.150.5

Interactive version: theaggregate.ai/benchmark?slug=glotocr-bench · How It Works · Data refreshed daily, snapshot 2026-10-07.