Google Unveils Three Variants of the Gemma Model

Google has released three variants based on the Gemma model:

– MedGemma, a medical model designed to run on a single GPU.
– SignGemma, a sign language translation model that supports multiple languages.
– DolphinGemma, a model capable of synthesizing dolphin sounds.

Among them, MedGemma offers two versions:

1. The 4B multimodal model: This version can handle tasks that involve a combination of images and text, having been pretrained on medical data including chest X-rays, dermatological images, ophthalmological images, and pathology slides.
2. The 27B textual reasoning model: With more robust text processing capabilities, this model is suitable for deeper reasoning tasks. It has been pretrained on medical text and image data, covering radiological imaging, pathology slides, ophthalmological images, and dermatological images.