MedGemma: 医療AI開発向けGoogleの最強オープンモデル MedGemma: Our most capable open models for health AI development

Google DeepMind and Google Research have introduced MedGemma, a family of open models tailored for medical AI development. Built on top of Gemma 3, the collection is designed to handle both medical imagery and clinical text, and crucially ships with open weights so developers can fine-tune and deploy it inside their own infrastructure.

The release includes three main components: multimodal 4B and 27B variants, and a standalone image encoder called MedSigLIP at roughly 400M parameters. MedSigLIP is a SigLIP-derived encoder tuned on chest X-rays, dermatology photos, fundus images, and histopathology slides, suitable for zero-shot classification and similarity search. The multimodal models target tasks such as radiology report generation, visual question answering on medical images, and broader clinical reasoning.

Google reports that MedGemma 27B reaches 87.7% on MedQA, competitive with much larger systems while running at roughly a tenth of the inference cost of DeepSeek R1. The 4B variant is positioned for resource-constrained settings such as on-device or on-premise hospital servers. Because the weights are fully open, hospitals and startups can run the models behind their own firewalls, an important factor for HIPAA compliance, reproducibility, and handling protected health information that cannot be sent to third-party APIs.

The broader context matters here. Healthcare has been a difficult fit for closed, API-only frontier models: data is siloed, regulated, and unevenly distributed across institutions and modalities. Competing efforts include Google's earlier Med-PaLM line, Microsoft's LLaVA-Med, and a growing body of work adapting general-purpose open backbones such as Meta's DINOv2 for pathology and radiology. MedGemma's pitch is essentially a ready-made open baseline that lowers the barrier to entry for clinical research labs and health-tech startups, who can then specialise the models with their own datasets.

Google is careful to note that MedGemma is not a clinical-grade product out of the box. Outputs require validation, fine-tuning, and appropriate regulatory clearance before any patient-facing use. Early adopters cited in the announcement include DeepHealth, which is exploring chest X-ray triage workflows, and Taiwan's Chang Gung Memorial Hospital, which has tested the models on Chinese-language medical literature. Other partners are reportedly evaluating the models for traditional Chinese medical text and histopathology subtyping.

Looking ahead, MedGemma's combination of a vision encoder plus paired generative models could encourage a more modular ecosystem, where institutions mix MedSigLIP-based retrieval with locally fine-tuned generation. Whether it becomes the default open foundation for medical AI will likely depend on how quickly the community produces robust evaluation benchmarks and post-training recipes — areas where the field still lags behind general-purpose LLM tooling.