OmniHandwritingOCR - Chinese Text: leaderboard
A subset of OmniHandwritingOCR (East China Normal University, 2026): 12,529 Chinese handwritten text images from CASIA-HWDB, filtered toward hard samples, and newly collected student writing. Systems transcribe each image to Markdown zero-shot, scored by 1-NED, one minus the normalized edit distance to the ground truth, in percent; higher is better.
Metric: 1-NED (%; one minus the normalized Levenshtein distance between the transcription and the ground truth, averaged over 12,529 Chinese handwritten text images (CASIA-HWDB and newly collected student writing); zero-shot Markdown transcription from one image with one task-neutral instruction for every system; task-specific normalization and tokenization, no semantic correction). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 79.91 | #738 |
| 2 | GPT-4o | 60.66 | #741 |
| 3 | Gemma 3 27B (IT) | 29.28 | #1289 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omnihandwritingocr-chinese-text · How It Works · Data refreshed daily, snapshot 2026-11-13.