At the 2025 Google I/O Developer Conference, Google officially released the lightweight multimodal model Gemma3n and announced the continued expansion of the Gemma model family, with the addition of MedGemma and SignGemma for medical and accessibility scenarios.
As a representative of the local AI trend, Gemma3n is designed specifically for low-performance devices such as smartphones, laptops, and tablets. It has the ability to process text, audio, images, and videos. According to Google, Gemma3n can run smoothly on devices with less than 2GB of running memory, demonstrating its efficient architecture advantage. The model was made available for developer preview on the day of release and shares the underlying architecture with Gemini Nano.
Google emphasized that the lightness and privacy advantages of Gemma3n make it an ideal choice for offline AI applications. Compared with large models that rely on the cloud, local AI models not only respond faster and cost less but also protect user data from being transmitted to remote servers, especially suitable for privacy-sensitive scenarios.
In the health field, Google's Health AI Developer Foundation introduced MedGemma—a set of open models targeting health-related text and images. It is reported that MedGemma has excellent multimodal analysis capabilities and supports developers in building more precise AI tools in scenarios such as medical image recognition and medical record text processing.
Additionally, Google previewed the upcoming SignGemma model, which is designed for sign language recognition. It can translate American Sign Language (ASL) into English text, aiming to help developers build more friendly communication tools for deaf and hard-of-hearing users. Google stated that this is "currently the most powerful sign language understanding model."
Despite concerns among some developers about commercial use due to the non-standard licensing terms of the Gemma model series, its broad functionality has received enthusiastic responses from the developer community. The download volume of the Gemma model series has exceeded tens of millions.
As the Gemma model family continues to expand, Google is promoting the application of open AI models in device-side, local privacy protection, healthcare, and accessibility fields.