MedGemma 1.5 4B

MedGemma 1.5 4B is Google's medical model built on Gemma 3, published on Hugging Face on 13 January 2026 as the only variant of this generation — there is no 27B counterpart, unlike in MedGemma 1. It keeps the architecture of the base model (decoder-only transformer, grouped-query attention, 128K tokens of input, 8,192 of output, images normalised to 896 × 896 and encoded as 256 tokens each), but both components are retrained on medical material: the SigLIP image encoder on de-identified X-rays, dermatology, ophthalmology and histopathology images, and the language part on medical text, question-answer pairs, electronic health records in the FHIR format, radiology in 2D and 3D, and laboratory reports. What this generation adds is mostly new kinds of input rather than higher scores. The model reads three-dimensional CT and MRI volumes, takes several patches of a whole histopathology slide at once, compares a chest X-ray with the patient's earlier images, marks anatomical structures and findings with bounding boxes, converts a scanned laboratory report into structured data, and reads text-based health records. The jumps on Google's own figures are where those abilities are new: description of a whole slide rises from 2.2 to 49.4 ROUGE against MedGemma 1 4B, anatomical box detection from 3.1 to 38.0 IoU, fundus classification from 64.9 to 76.8 per cent accuracy, and question answering over health records from 67.6 to 89.6. The catalogue notes this honestly: the progress is not uniform. On the SLAKE radiology question set the new model falls from 72.3 to 59.7 tokenised F1 — Google explains that it was tuned less for that question format — and on PubMedQA it drops from 73.4 to 68.2. Chest X-ray report generation, at 27.2 RadGraph F1, is below the specially tuned MedGemma 1 4B at 30.3. On medical text the model is clearly better than its predecessor (MedQA 69.1 against 64.4) yet still well behind the 27B model from the previous generation at 85.3. The weights are downloadable after accepting Google's Health AI Developer Foundations terms, which permit commercial use but bind the user to a restricted-use policy. Google states clearly that this is not a clinical-grade product: it is a starting point for developers, to be fine-tuned and validated before any use that touches a patient.

▸Videos

No videos yet.

▸Reviews