Google’s MedGemma: A Stethoscope for the Algorithm Age

Google has a habit of naming its creations like Pokémon evolutions. Enter Med-Gemini (the research flagship) and MedGemma (the open-source cousin you can actually download). Both are medical large language models, and both are Google’s way of saying: “Yes, AI can memorise your anatomy textbook and squint at your X-ray, but please, don’t sue us just yet.”

Med-Gemini is DeepMind’s high-end research line, built on Gemini. It reads text, looks at scans, keeps its memory long enough to parse an entire patient record without keeling over, and, on paper, it’s smashing benchmarks. Google claims it outperformed GPT-4 on every comparable medical test, even hitting 91% on USMLE-style exams. Impressive, though it did once invent a body part called the “basilar ganglia”, which is not a thing! I think?

MedGemma, by contrast, is the pragmatic sibling. Built on Gemma 3, it comes in 4B and 27B flavours, and Google has thrown it open for hospitals, startups and hobbyists to fine-tune. The big selling point: it can run locally or on Vertex AI, so your sensitive patient data doesn’t have to ping across to Mountain View. For regulators and compliance officers, that’s less “Wild West” and more “NHS-friendly.”

The implications are clear. We’re on the cusp of seeing agentic copilots that sift through charts, pull out the relevant labs, and draft a plan while flagging uncertainty. Radiology scribes that give first-pass reads of X-rays or derm photos will appear in clinics, producing structured drafts for humans to sanity-check. And because MedGemma is open, expect a cottage industry of bespoke models tuned to local acronyms, national guidelines, and niche specialities.

Of course, the uncomfortable bit is still there: automation bias. When a model sounds confident, humans tend to nod along, even when it’s wrong. A hallucination in a meme generator is funny. A hallucination in an oncology consult is malpractice. Until these systems come with iron-clad traceability, uncertainty scores, and regulatory sign-off, the real doctor is still the one in charge.

MedGemma might look like just another AI release, but it represents a subtle power shift. By making a capable medical model open and adaptable, Google is democratising access in a field usually locked behind proprietary walls. Expect national health systems, private clinics, and startups to carve MedGemma into shapes that fit their workflows. Some of those shapes will save time, money, and maybe lives. Others will almost certainly end up in the “what were they thinking” pile.

The bottom line is simple: Med-Gemini is the glossy research prototype, MedGemma is the open toolkit you can tinker with today, and both show where medicine is heading, algorithms as assistants, not overlords. The basilar ganglia may not exist, but liability insurance definitely should.