WildHandBench: leaderboard
WildHandBench (2026) tests handwritten document understanding on 500 real-world handwritten documents, free text, tables and formulas in four languages and nine capture scenarios, with a calibrated human baseline. The overall score in percent averages text accuracy (100 minus normalized edit distance), table TEDS and formula CDM; higher is better.
Metric: Overall score (%; mean of 100 minus the text normalized edit distance, the table TEDS and the formula CDM over 500 real handwritten documents (free text, tables and formulas) in four languages and nine real-world scenarios such as notes, forms, whiteboards and exam sheets). Source: arxiv.org. Saturation forecast: Around April 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 71.85 | #94 |
| 2 | Gemini 3.5 Flash | 68.25 | #101 |
| 3 | Kimi K3 | 67.81 | #54 |
| 4 | Qwen 3.5 Plus | 65.55 | #232 |
| 5 | Qwen 3 VL 235B A22B (Thinking) | 62.19 | #451 |
| 6 | Claude Opus 4.8 | 57.7 | #56 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=wildhandbench · How It Works · Data refreshed daily, snapshot 2026-11-13.