Google unveils ‘most intelligent’ open AI models that run offline on phones

Google’s Gemma 4 models bring advanced AI capabilities to smartphones and edge devices, enabling offline performance, enhanced privacy, and flexible deployment while supporting developers with open-source access and scalable model options.

Google has unveiled Gemma 4, a new family of open AI models designed to run directly on devices, including smartphones, without requiring constant internet connectivity. The company positions the models as among the most capable in the open ecosystem, combining strong performance with the flexibility of local deployment.

Announced by Demis Hassabis, the release marks a shift toward making powerful AI tools more accessible beyond cloud-based environments. By enabling on-device processing, Gemma 4 aims to deliver faster responses and improved privacy, as user data remains stored locally rather than being transmitted to external servers.

Designed for performance across devices

The Gemma 4 lineup includes multiple model sizes, allowing developers to select configurations suited to different computing environments. Smaller models are optimised for mobile and edge devices, while larger variants deliver more advanced reasoning and computational capabilities.

Google says the models support complex tasks such as coding, multi-step reasoning, and agent-based workflows. These features allow users to assign tasks to AI systems that can execute them with minimal manual intervention, expanding the scope of practical applications.

The models are released under an open-source licence, giving developers the ability to modify and deploy them freely. This approach is intended to foster innovation and broaden participation in AI development, particularly among startups and independent developers.

Advancing offline AI capabilities

A key highlight of Gemma 4 is its ability to operate offline on everyday devices, including Android smartphones. This capability could significantly expand access to AI, especially in regions with limited connectivity, while also addressing growing concerns around data privacy.

In addition to text processing, the models support multiple input formats such as images and video, with select variants also handling audio. Enhanced context windows enable the models to process longer inputs, including detailed documents and codebases, within a single prompt.

The launch also builds on the growing ecosystem around Gemma, which has already seen widespread adoption among developers. By combining open access with strong performance, Google is positioning Gemma 4 as a key step in bringing advanced AI closer to everyday users.

With this release, Google is reinforcing its focus on open and accessible AI, as competition intensifies to deliver powerful yet practical solutions that can operate seamlessly across devices.