Google Releases MedGemma 1.5: An Open Source AI That Can Read 3D Imaging

— Sundar Pichai, CEO of Google

Google just released MedGemma 1.5, an updated version of their open medical AI model. It's free for research and commercial use.

You can run MedGemma 1.5 4B using LM Studio. Since it’s a Google product, Gemini can easily walk you through setup for your specific system. (I wouldn’t suggest using it for patient images at this point, but if you’ve had any of your own imaging done, it could be interesting to try.)

MedGemma 1.5 4B is an updated open medical AI model. “Open” means developers can download it, modify it, and build products on top of it. The “4B” refers to 4 billion parameters, which means it is small enough to run efficiently while still being capable.

Key improvements:

• 3D medical imaging interpretation. MedGemma 1.5 can now natively interpret CT scans, MRIs, and whole-slide pathology images. Previous versions only handled 2D images like chest X-rays.

• Longitudinal chest X-ray comparison. The model can compare X-rays taken at different times and describe changes. Accuracy improved 5% over the previous version.

• Anatomical localization. Identification of anatomical structures in chest X-rays improved by 35%.

• Better text performance. Medical question-answering accuracy improved 5-22% depending on the task.

Until now, most medical AI models could only handle 2D images. A chest X-ray, a skin lesion photo, a single slice.

MedGemma 1.5 can now accept multiple CT or MRI slices as input and reason across them. Google claims this is the first open multimodal model that can do this while also handling text, 2D images, and general reasoning.

Google explicitly states this model is not ready for clinical use without significant validation. The baseline capabilities are imperfect. But this is the foundation developers will build on.

A 4B parameter model can run on hardware you might already own.

At full precision, MedGemma 1.5 4B needs roughly 8 GB of GPU memory. An NVIDIA RTX 3060 gaming GPU has 12 GB. A MacBook Pro with 16 GB of unified memory could run this locally.

For comparison, GPT-4 is estimated at 1.8 trillion parameters. That requires specialized data center hardware with hundreds of gigabytes of GPU memory. You can’t run it locally. You access it through the cloud, and OpenAI controls the infrastructure. MedGemma 1.5 4B is 450 times smaller.

MedGemma 1.5 4B is “small enough to run offline.” A hospital IT department, a health tech startup, or an individual developer with a decent laptop could experiment without cloud costs. There’s a tradeoff: the 4B model is less capable than Google’s larger 27B version for complex tasks. But the direction is clear. Medical AI is becoming something that runs on local hardware, not just cloud data centers.

Models are getting smaller and more capable, with more intelligence per gigabite. Hardware requirements drop. What required a data center two years ago runs on a laptop today.

MedGemma 1.5 isn’t going to read your CT scans next week. But the developer building your hospital’s next radiology tool might start with this model as their foundation. The tools that eventually reach your workflow will have roots in releases like this one.