SeaEval - Multilingual Reasoning - CMMLU (Zero-Shot): leaderboard
SeaEval zero-shot CMMLU evaluation measuring multilingual subject knowledge and exam-style reasoning accuracy.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B Instruct | 83.44 | #454 |
| 2 | Qwen 2 72B Instruct | 82.94 | #576 |
| 3 | Qwen 2.5 32B Instruct | 82.73 | #498 |
| 4 | Qwen 2.5 14B Instruct | 78.08 | #647 |
| 5 | Qwen 2 7B Instruct | 77.28 | #991 |
| 6 | SeaLLMs-v3-7B-Chat | 76.84 | #915 |
| 7 | Qwen 2.5 7B Instruct | 74.87 | #862 |
| 8 | Llama 3.1 70B Instruct | 68.15 | #561 |
| 9 | Qwen 2.5 3B Instruct | 66.21 | #1162 |
| 10 | Llama 3 70B Instruct | 64.95 | #634 |
| 11 | Sailor2-8B-Chat | 64.17 | #782 |
| 12 | Qwen 2.5 1.5B Instruct | 59.76 | #1407 |
| 13 | gemma2-9B-cpt-sea-lionv3-instruct | 57.96 | #705 |
| 14 | Gemma 2 9B (IT) | 57 | #788 |
| 15 | Llama 3.1 8B Instruct | 51.83 | #1035 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=seaeval-multilingual-reasoning-cmmlu-zero-shot · How It Works · Data refreshed daily, snapshot 2026-10-20.