MedMosaic: leaderboard
MedMosaic (Rajgarhia et al., 2026) is a medical audio question-answering benchmark over heart, lung and cough sounds, short and long clinical conversations, conversations with inserted clinical sounds, spoken questions and multi-turn chains. The board weights each category score by its number of questions; multiple-choice categories are accuracy and open-ended ones a GPT-5.1 judge score, in percent (higher is better).
Metric: Weighted accuracy (%; category scores weighted by question count over 46,821 scored questions, the 60 multi-turn chains counted as 180 turns; open-ended categories scored by a GPT-5.1 judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 68.1 | #262 |
| 2 | Gemini 2.5 Flash | 60.5 | #452 |
| 3 | Qwen2.5-Omni-7B | 42.8 | #1797 |
| 4 | Gemma 3n E4B (IT) | 42.1 | #2220 |
| 5 | Phi-4 Multimodal Instruct | 37.3 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medmosaic · How It Works · Data refreshed daily, snapshot 2026-11-12.