BanglaWild (Bengali Prompt): leaderboard
BanglaWild (2026) tests Bengali scene text recognition on 1,014 in-the-wild mobile-phone photographs of walls, signboards, vehicles, banners and product packaging, scored against verbatim human transcriptions. Character error rate over Unicode code points after NFC normalization, times 100; lower is better, and it can exceed 100 when a model emits long off-target output. Vision-language models get the same transcribe-only instruction written in Bengali; conventional OCR engines run without a prompt.
Metric: Character error rate (CER x100: character edit distance over Unicode code points after NFC normalization divided by the reference length, mean over the 1,014-image test split; not capped, so it can exceed 100 when a model emits long off-target output; zero-shot, temperature 0 or greedy, 150-token cap; the same instruction in Bengali script, prompt P2, image only). Source: arxiv.org. Saturation forecast: Around April 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 15.6 | #452 |
| 2 | Claude Sonnet 4.6 | 18.46 | #158 |
| 3 | Qwen 3.5 35B A3B | 20.23 | #480 |
| 4 | Qwen 3.5 9B | 20.51 | #803 |
| 5 | Gemini 2.5 Pro | 21.65 | #262 |
| 6 | Qwen 3.5 27B | 23.45 | #358 |
| 7 | Llama 4 Maverick | 37.98 | #1105 |
| 8 | GPT-5.4 Mini | 56.83 | #389 |
| 9 | Gemma 3 12B (IT) | 70.76 | #1849 |
| 10 | Claude Haiku 4.5 | 82.17 | #568 |
| 11 | Gemma 3 4B (IT) | 119.64 | #3848 |
| 12 | GPT-5.4 Nano | 192.8 | #633 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=banglawild-bengali-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.