UrduMMLU (Urdu Prompt): leaderboard
Metric: Overall exact-match accuracy (%) on UrduMMLU's Urdu multiple-choice questions from Pakistani MCQ banks and public examinations, zero-shot, answer key parsed from the generation (unparsable, malformed or error outputs count as wrong); temperature 0 where available, 4,096 output tokens, Urdu instruction prompt. Source: arxiv.org. Saturation forecast: Around December 2026. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.5 Flash | 90.45 |
| 2 | Gemini 3.1 Flash Lite | 84.81 |
| 3 | GPT-5.4 | 84.32 |
| 4 | Claude Sonnet 4.6 | 83.07 |
| 5 | DeepSeek V4 Flash | 81.33 |
| 6 | Gemma 4 31B (IT) | 76.49 |
| 7 | Llama 4 Maverick Instruct | 75.93 |
| 8 | Qwen 3.6 35B A3B | 75.25 |
| 9 | GPT-5.4 Mini | 73.63 |
| 10 | Claude Haiku 4.5 | 72.52 |
| 11 | Gemma 4 26B A4B (IT) | 70.32 |
| 12 | Qwen 3.6 27B | 69.81 |
| 13 | Llama 4 Scout Instruct | 68.28 |
| 14 | Llama 3.3 70B Instruct | 67.1 |
| 15 | Ministral-3-8B-Instruct-2512 | 57.08 |
Interactive version: theaggregate.ai/benchmark?slug=urdummlu-urdu-prompt · How It Works · Data refreshed daily, snapshot 2026-09-29.