Devanagari OCR Stress Test - Real Scans: leaderboard
Devanagari OCR stress test (Aditya Pratap Singh, 2026) comparing classical OCR engines, open and specialised OCR vision-language models and frontier multimodal APIs on Hindi text. This set is 300 word- and short-phrase images of historical typeset Devanagari print from the Sanskrit-OCR-Typed corpus. Transcriptions are NFC-normalised and scored with chrF++ (character n-gram F-score with word bigrams, 0-100); higher is better.
Metric: chrF++ (0-100), 300 real printed word and short-phrase images from Sanskrit-OCR-Typed, NFC-normalised. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 86.3 | #452 |
| 2 | Claude Opus 4.7 | 82.2 | #75 |
| 3 | Qwen 3 VL 8B Instruct | 75.2 | #982 |
| 4 | GPT-5.5 | 58.5 | #48 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=devanagari-ocr-stress-test-real-scans · How It Works · Data refreshed daily, snapshot 2026-11-12.