MedGemma-4B: benchmark results
Provider: Google. Released 2025-05-20.
Aggregate Rating 1485 ± 18, rank #2444 of 10800 rated models, from 85 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LoMeVQA - Progress Classification | 52.6 | Accuracy (%; three-option classification of disease progress | 93.3 |
| MI-CXR - Ordered Event Localization | 30.1 | Temporal Event Localization, ordered emergence-then-resoluti | 92.3 |
| EPAG - Top-k Diagnosis Accuracy | 82.31 | Disease Diagnosis Accuracy, Gold Anywhere in Ranked List (%) | 85 |
| Large Language Models Lack Temporal Awareness | 60.1 | Accuracy (self-reported) | 84.6 |
| LoMeVQA - Progress Description | 20.95 | F1-RadGraph clinical efficacy (%; free-text description of t | 80 |
| MedErrBench (Arabic) - Error Correction | 43.8 | Correction ROUGE-1 (%): unigram overlap of the corrected sen | 76.9 |
| MedErrBench (English) - Error Correction | 51.1 | Correction ROUGE-1 (%): unigram overlap of the corrected sen | 76.9 |
| MedQ-Deg - Treatment Reasoning | 25 | Accuracy (%) on the Treatment questions of MedQ-Deg (capabil | 75.6 |
| MedErrBench (Arabic) - Error Localization | 43.3 | Localization accuracy (%): naming the sentence that contains | 75 |
| PSF-Med - Open Medical - PadChest | 13.4 | Paraphrase Flip Rate (%) | 75 |
| MedErrBench (Chinese) - Error Correction | 54.9 | Correction ROUGE-1 (%): unigram overlap of the corrected sen | 69.2 |
| MedErrBench (Chinese) - Error Localization | 50 | Localization accuracy (%): naming the sentence that contains | 69.2 |
Interactive version: theaggregate.ai/model?slug=medgemma-4b · How It Works · Data refreshed daily, snapshot 2026-11-13.