Pixel 10 Tensor G5 Runs Gemma 4 E2B Natively Offline

At Google I/O Connect India, Google shifted its mobile hardware strategy toward local inference, announcing that the Pixel 10 device family will run the new Gemma 4 E2B model entirely offline. The release is powered by the Tensor G5, Google’s first fully in-house System on Chip manufactured on a 3nm process.

For developers building mobile applications, the update provides a standardized path to integrate local multimodal AI features without requiring internet connectivity or external API calls.

Custom Silicon and the Tensor G5

The Tensor G5 represents a departure from Google’s previous Samsung-based designs, focusing heavily on thermal efficiency and low-latency transformer inference. The chip includes an upgraded Tensor Processing Unit optimized for INT8 and FP16 hardware acceleration.

To support the “100% private” execution claims, the Tensor G5 incorporates a new Private Compute Core enclave. This hardware-level isolation prevents AI model weights and active session data from leaving the device’s volatile memory during active inference, addressing data security constraints for enterprise and government use cases.

Gemma 4 E2B Specifications

Gemma 4 E2B (Edge-to-Brain) is a lightweight, multimodal variant of the Gemma 4 family distilled specifically for mobile deployment. It processes text, image, and audio inputs locally.

Demonstrated performance metrics include:

• Image Recognition: 30ms latency for instantaneous object and scene description via the device camera.
• Translation: Real-time speech-to-speech translation supporting 45 languages, functional in Airplane Mode.
• Local Synthesis: High-fidelity local audio synthesis, with specific optimizations for 12 Indian languages.
• Document Summarization: Offline processing of local files via the native Personal Assistant mode.

While security analysts praised the isolation architecture, developers have noted that storing high-precision multimodal weights may pressure storage limits on the base 128GB Pixel 10 models.

Developer Integration and Availability

The offline features are confirmed for the Pixel 10, Pixel 10 Pro, and Pixel 10 Pro XL. Third-party access is managed through an updated version of AICore, which handles the lifecycle of the model on the device.

Developers can access the Gemma 4 E2B model for local application features through the Google Play Services SDK. The model weights were made available to early access partners via the Google AI Edge platform starting July 13, 2026. If you plan to run Gemma 4 on-device for your own Android applications, evaluating the memory footprint against the device’s storage limits will be a necessary step in the deployment process.