CureMed-Bench: leaderboard
Open-ended multilingual medical reasoning from CureMed-Bench (Onyame et al., 2026), built from MedlinePlus content and reviewed by medical experts and native speakers in 13 languages, including Amharic, Hausa, Swahili and Yoruba. Logical accuracy in percent: the share of final answers a GPT-4o judge matches to the reference answer, pooled over 1,431 held-out questions, zero-shot; higher is better.
Metric: Logical accuracy (%; share of final answers a GPT-4o judge matches to the reference answer, pooled over 1,431 held-out open-ended medical questions in 13 languages; zero-shot). Source: arxiv.org. Saturation forecast: Around February 2027. 49 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 80.57 | #349 |
| 2 | Claude 3 Haiku | 73.31 | #2629 |
| 3 | GPT-5 Nano | 73.24 | #888 |
| 4 | Llama 3.3 70B Instruct | 60.8 | #1353 |
| 5 | Qwen 2.5 72B Instruct | 58.8 | #1093 |
| 6 | Gemini 2.5 Flash | 54.79 | #452 |
| 7 | Llama 3.1 70B | 54.65 | #1493 |
| 8 | Qwen 2.5 32B Instruct | 49.69 | #1216 |
| 9 | Qwen 3 32B | 48.57 | #960 |
| 10 | Ministral 8B | 42.87 | #4511 |
| 11 | Qwen 2.5 14B Instruct | 41.79 | #1745 |
| 12 | DeepSeek R1 Distill Qwen 32B | 40.27 | #1780 |
| 13 | Qwen 3 14B | 38.62 | #1235 |
| 14 | Gemma 2 9B | 36.97 | #2881 |
| 15 | Llama 3 8B UltraMedical | 35.29 | #6117 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=curemed-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.