Google Gemma AI Model Now Runs on Phones

Google Expands its Gemma Family of Open AI Models

The range of Google’s “open” artificial intelligence models, known as Gemma, is currently expanding.

At Google I/O 2025 on Tuesday, Google unveiled Gemma 3n, a model specifically engineered for seamless operation on smartphones, laptops, and tablet computers. This model is available in preview as of Tuesday and is capable of processing audio, text, images, and video content, as stated by Google.

The Rise of Efficient, Offline AI

AI models that are sufficiently efficient to function offline, without reliance on cloud-based computing, have been gaining prominence within the AI community. These models offer cost advantages compared to larger counterparts and enhance user privacy by eliminating the necessity of data transmission to remote servers.

During a keynote presentation at I/O, Gus Martins, Gemma Product Manager, highlighted that Gemma 3n can operate effectively on devices with less than 2GB of RAM. He further noted, “Gemma 3n utilizes the same architecture as Gemini Nano and is designed for exceptional performance.”

New Models: MedGemma and SignGemma

Alongside Gemma 3n, Google is also releasing MedGemma through its Health AI Developer Foundations program. The company asserts that MedGemma represents its most advanced open model for the analysis of health-related text and images.

“MedGemma is our collection of open models designed for multimodal understanding of health-related text and images,” Martins explained. “It excels in a variety of image and text applications, allowing developers to customize the models for their specific health applications.”

Furthermore, Google is introducing SignGemma, an open model intended to translate sign language into spoken-language text. Google anticipates that SignGemma will empower developers to create innovative applications and integrations for individuals who are deaf or hard of hearing.

“SignGemma is a new family of models trained to translate sign language into spoken text, with particular proficiency in American Sign Language and English,” Martins said. “It represents the most capable sign language understanding model to date, and we are eager to see developers and the deaf and hard-of-hearing communities build upon this foundation.”

Licensing Considerations and Adoption

It is important to note that Gemma has faced some criticism regarding its custom licensing terms. Certain developers have expressed concerns that these terms introduce commercial risks when utilizing the models. Despite this, Gemma models have been downloaded collectively tens of millions of times, demonstrating significant developer interest.

Updated 2:40 p.m. Pacific: Additional quotes from Gemma Product Manager Gus Martins have been included.