ParsiEval: leaderboard
364 Persian multiple-choice questions on history, literature, general knowledge and science, testing comprehension, fact recall and inference. Accuracy, higher is better. Source: github.com/mshojaei77/ParsiEval.
Metric: Accuracy (%, 364 questions). Source: github.com. Saturation forecast: Estimated already saturated. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 84.62 | #159 |
| 2 | GPT-4o | 79.12 | #357 |
| 3 | GPT-5 Mini | 77.75 | #193 |
| 4 | Kimi K2 | 77.2 | #252 |
| 5 | Llama 4 Maverick Instruct | 75.27 | #467 |
| 6 | DeepSeek V3 (0324) | 74.73 | #349 |
| 7 | Llama 3.3 70B | 73.35 | #578 |
| 8 | Llama 4 Scout Instruct | 71.43 | #547 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 70.6 | #315 |
| 10 | GPT-OSS-120B | 70.05 | #346 |
| 11 | Gemma 3 27B (IT) | 67.03 | #528 |
| 12 | Qwen 3 30B A3B (Thinking) | 65.66 | #496 (Qwen 3 30B A3B) |
| 13 | GPT-4.1 Nano | 63.74 | #730 |
| 14 | GPT-OSS-20B | 63.74 | #523 |
| 15 | Gemma 2 9B (IT) | 59.62 | #795 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=parsieval · How It Works · Data refreshed daily, snapshot 2026-11-05.