MedGemma 1.5 4B: benchmark results
Provider: Google. Released 2026-01-13. Access: Open.
Aggregate Rating 1497 ± 11, rank #788 of 1635 rated models, from 102 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MC-CXR - Image-Only Recognition | 31.7 | Image-only recognition accuracy (%; the image alone, no cont | 88.9 |
| MC-CXR - Misleading Preliminary Note | 12.1 | Accuracy under misleading context (%; the image plus a preli | 88.9 |
| Holtercare-Bench - Presence (Text) | 62.42 | Accuracy (%; multiple-choice closed QA on whether a given rh | 83.3 |
| PSF-Med - Open Medical - MIMIC-CXR | 7.4 | Paraphrase Flip Rate (%) | 80 |
| MC-CXR - Misleading Overlay | 19.6 | Accuracy under misleading context (%; the image plus a PACS- | 77.8 |
| Holtercare-Bench - Diagnosis (Text) | 33.46 | Accuracy (%; multiple-choice closed QA on the overall diagno | 75 |
| DeepTumorVQA - Medical Reasoning | 29 | Accuracy (%) on the Medical-reasoning-stage questions (two t | 63.6 |
| Holtercare-Bench - Heart Rate Extremum Timing (Text) | 34.71 | Accuracy (%; multiple-choice closed QA on the millisecond ti | 62.5 |
| MedFlowBench (Radiology) - Brain MRI Task | 45 | Diagnosis accuracy (0-1, shown times 100) on UCSF-PDGM multi | 61.5 |
| Oncology Report-Derived VQA - Liver CT Report-Derived | 47.8 | Accuracy (%; zero-shot multiple-choice VQA with 2-5 options | 60 |
| Holtercare-Bench - Event Timing (Text) | 30.79 | Accuracy (%; multiple-choice closed QA on the millisecond ti | 58.3 |
| MC-CXR - Misleading Prior Report | 13.3 | Accuracy under misleading context (%; the image plus a prior | 55.6 |
Interactive version: theaggregate.ai/model?slug=medgemma-1-5-4b · How It Works · Data refreshed daily, snapshot 2026-11-05.