GMAI-MMBench: leaderboard

Each vision-language model answers zero-shot questions about medical images (38 imaging modalities, from public datasets and hospitals) on the held-out Test split; the score is the percentage of the split's single-choice questions answered correctly, as the authors report it. The Test split holds 21,281 questions; its multiple-choice items are scored separately and are not in this number. Unlike text-only medical exams such as MedQA, every question needs the image, and many point at a box, mask or contour in it.

Metric: Accuracy on single-choice questions (%). Source: github.com. Saturation forecast: Around February 2027. 24 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1MedDr-40B43.69
2DeepSeek-VL-7B-Chat43.43#3480
3MiniCPM-V-242.54#7982
4Qwen-VL-Max42.16#1332
5LLaVA-internLM2-7B40.45
6InternVL-Chat-V1.539.73#3908
7OmniLMM-12B39.3#8047
8InternLM-XComposer2-VL-7B39.2
9TransCore-M38.7#8070
10LLaVA-v1.5-7B37.96#7933
11Emu2_chat Chat37.59#7446
12Qwen-VL-Chat36.96#8089
13mPLUG-Owl236.21#8207
14CogVLM-17B-Chat36.08#8696
15Yi-VL-6B34.31#8065

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=gmai-mmbench · How It Works · Data refreshed daily, snapshot 2026-11-17.