BankMathBench (Korean) - Intermediate: leaderboard
Metric: Exact-match accuracy (%) on the 426 Korean intermediate (comparison of two products by interest rate or total payout) test questions of BankMathBench, banking calculations on deposits, savings and loans; zero-shot, one instruction format for all models, equivalent number and currency formats accepted; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 10.7 | #237 |
| 2 | Claude 3.5 Sonnet (20240620) | 10.3 | #293 |
| 3 | GPT-4o (2024-05-13) | 7.9 | #316 |
| 4 | Qwen 2.5 7B Instruct | 3 | #846 |
| 5 | Qwen 3 8B | 2.8 | #667 |
| 6 | Qwen 3 4B | 2.6 | #823 |
| 7 | DeepSeek-R1-Distill-Qwen-7B | 1.6 | #1400 |
| 8 | Qwen2.5-Math-7B-Instruct | 0.9 | #1422 |
| 9 | Llama 3 8B Instruct | 0.7 | #1115 |
| 10 | Llama 3.2 3B Instruct | 0.5 | #1321 |
| 11 | Phi-3.5-mini-instruct | 0.5 | #1148 |
| 12 | Qwen 2.5 3B Instruct | 0.2 | #1138 |
| 13 | Llama 3.2 1B Instruct | 0 | #1518 |
| 14 | Qwen 2.5 1.5B Instruct | 0 | #1388 |
| 15 | DeepSeek R1 Distill Qwen 1.5B | 0 | #1542 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=bankmathbench-korean-intermediate · How It Works · Data refreshed daily, snapshot 2026-10-11.