JMed48k (With Images): leaderboard
Metric: Accuracy (%) on the 2,579 scored JMed48k-Eval items that include images, unweighted macro average over the 11 professions, official MHLW gold answers, single-shot Japanese prompts at temperature 0 with reasoning disabled where the model has a switch, JSON-constrained answers; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 79.7 |
| 2 | GPT-5 | 78.8 |
| 3 | GPT-5 Mini | 70.6 |
| 4 | Gemini 2.5 Flash (Non-reasoning) | 66.7 |
| 5 | Qwen 3.5 397B A17B (Non-reasoning) | 63.6 |
| 6 | Gemma 4 31B (IT) | 61.6 |
| 7 | Claude Sonnet 4 | 57.7 |
| 8 | Qwen 3.5 27B (Non-reasoning) | 57.6 |
| 9 | Llama 4 Maverick | 53.8 |
| 10 | Grok 4.20 | 53.6 |
| 11 | Lingshu-32B | 45 |
| 12 | Qwen 3.5 9B (Non-reasoning) | 41.8 |
| 13 | MedGemma-27B-IT | 34.2 |
| 14 | MedGemma-4B-IT | 23.6 |
Interactive version: theaggregate.ai/benchmark?slug=jmed48k-with-images · How It Works · Data refreshed daily, snapshot 2026-10-07.