It isn't Apache weights and it isn't a diagnostic device. Google Research posted MedGemma 1.5 4B on 13 January 2026 as a gated Hugging Face checkpoint under Health AI Developer Foundations terms. The card says outputs are not intended to inform clinical diagnosis.

What it does

The 13 January 2026 Google Research blog, by Daniel Golden and Fereshteh Mahvar, describes MedGemma 1.5 4B as an update that adapts MedGemma for high-dimensional imaging (CT, MRI, histopathology), longitudinal chest X-ray review, anatomical localization on chest X-rays, and structured extraction from medical lab reports. The official model card, last updated 21 April 2026 UTC, repeats those jobs and adds electronic health record text understanding. It states MedGemma 1.5 4B is an updated version of MedGemma 1 4B, built on Gemma 3 with a SigLIP image encoder pre-trained on de-identified medical images.

The model card's intended-use section says MedGemma is a starting point for developers in life sciences and healthcare to train, adapt, and change for a specific task.

Why it is moving now

The model card release notes date the 1.5 update to 13 January 2026: "Updated to version 1.5 with improved medical reasoning, medical records interpretation and medical image interpretation." That is the event. CT/MRI volume slices and whole-slide patches are why a hospital ML team would reopen a 2025 MedGemma 1 4B eval.

Google's blog claims internal benchmark lifts versus MedGemma 1 4B, including figures on CT and MRI finding classification. Those numbers are Google's, on internal suites. The model card states evaluations use both open benchmark datasets and internally curated datasets, and that developers are expected to fine-tune for their use case. Treat the blog deltas as vendor-reported until you rerun on your scanners, sites, and labels.

License + activity + maturity (demo | usable | production-shaped)

Weights and model derivatives: Health AI Developer Foundations Terms of Use (last modified 15 November 2024), which the model card names as the governing license. HAI-DEF recipe source, inference, training, and utility code are licensed under Apache 2.0, the terms state. Distribution of HAI-DEF or model derivatives requires passing through Section 3.2 use restrictions, giving recipients a copy of the Agreement, seeking health-regulatory authorization when applicable, marking modified files, and (except hosted service) shipping a Notice file pointing at the terms URL.

Section 3.2 bars uses in the Prohibited Use Policy, any use that could cause a health regulator to deem Google a medical-device "manufacturer," and illegal uses. Section 4.3 states HAI-DEF services and outputs are provided "AS IS" and that Google will not be deemed to furnish medical advice by providing the services. Maturity: usable as a research and product-prototyping checkpoint for teams that can accept HAI-DEF, run local or Vertex jobs, and keep a clinician out of the loop until a validated device pathway exists. Demo if you only click the Hugging Face widget. Production-shaped for clinical diagnosis: no. The model card: outputs "are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications."

When to pick it vs the default alternative

Pick MedGemma 1.5 4B when you need a 4B-class medical multimodal starting point that the card says can take CT/MRI volumes, histopathology patches, or lab-report images, and you can live under HAI-DEF rather than MIT or Apache weights.

Pick MedGemma 1 27B (text) when the job is long medical text and the 4B multimodal path is the wrong size. The January blog still points developers to the 27B for more complex text.

Pick a closed medical API (vendor EHR copilots, radiology SaaS) when you need a business associate agreement, a 510(k) story, or a support contract, not a fine-tune.

Pick a general Gemma 3 or other open VLM only for non-clinical prototyping. The card's point is medical continued pre-training. Whether 1.5 4B beats those general models on your holdout is UNKNOWN until you measure.

Don't pick it as a drop-in scribe that writes into the chart. Intended use forbids direct clinical practice outputs.

Failure modes

Google-reported accuracy deltas will not transfer to your scanner vendor, slice thickness, or language. The model card limitations still say multimodal capabilities were primarily evaluated on single-image tasks, that MedGemma has not been evaluated for comprehension of multiple images or optimized for multi-turn use, even as 1.5 advertises volumes, patches, and longitudinal X-rays. That tension is in Google's own card. Validate the new modalities yourself.

HAI-DEF can be terminated if Google reasonably determines a regulator may treat Google as a device manufacturer because of your use. You must then delete copies. Hosted redistribution is Distribution under the terms. Fine-tunes are model derivatives and carry the restrictions.

Outputs require independent verification. Child-safety and content-safety tests described on the Hugging Face card were Google's, without safety filters, and are not a hospital IRB result.

PHI in prompts and in fine-tune sets remains your HIPAA/GDPR problem. The checkpoint does not absorb it.

Links

Whether MedGemma 1.5 4B is accurate on your CT/MRI protocol is UNKNOWN.