medgemma-4b-it

MedGemma 4B IT (google/medgemma-4b-it) is Google's instruction-tuned, open-weight medical foundation model built on Gemma 3. It accepts text and medical images as input and produces text output, targeting developers building healthcare AI applications such as clinical decision support, medical document understanding, radiology VQA, and patient-facing assistants. The model combines a SigLIP-based MedSigLIP vision encoder—pre-trained on de-identified radiology, histopathology, dermatology, and ophthalmology data—with a 4-billion-parameter decoder-only Transformer using grouped-query attention (GQA) and a 128K-token context window.

MedGemma 4B IT was released on May 20, 2025 (version 1.0.1 with a multimodal bug fix on July 9, 2025) under the Health AI Developer Foundations program. It substantially outperforms the base Gemma 3 4B on medical benchmarks: 64.4% on MedQA (vs. 50.7%), 55.7% on MedMCQA, 73.4% on PubMedQA, and 70.0% on MMLU-Med. On imaging tasks it achieves 88.9 macro-F1 on MIMIC-CXR top-5 classification, 72.3 tokenized F1 on SLAKE radiology VQA, and strong results on dermatology, histopathology, and fundus classification benchmarks.

On general non-medical benchmarks, specialization incurs only minor trade-offs: 39.1% on MMLU-Pro and 47.3% on MMMU (validation) versus 43.6% and 48.8% for Gemma 3 4B. MedGemma is intended as a developer starting point, not for direct clinical diagnosis or treatment decisions without domain-specific validation, fine-tuning, and independent verification of all outputs.