Google releases Gemma 4 E2B, optimized to run natively on Pixel 10's Tensor G5 TPU

Google announced Gemma 4 E2B for TPU today, a variant of its open-source Gemma 4 model designed to run natively on the Tensor Processing Unit in Pixel 10 devices. The announcement came at I/O Connect India, following a similar satellite event in Berlin last week.

Model specifications

Gemma 4 E2B runs on the Tensor G5's TPU and is supported on four devices: Pixel 10, 10 Pro, 10 Pro XL, and 10 Pro Fold. Google describes it as "state-of-the-art, powerful, yet remarkably lightweight," though specific parameter counts and benchmark scores were not disclosed.

The model is based on Gemma 4, which Google first introduced in April as the foundation for the upcoming Gemini Nano 4. Gemma is Google's series of open models designed for on-device execution.

Multimodal capabilities

Gemma 4 E2B supports three fully offline modes:

• AI Chat: On-device conversations with no internet connection required
• Ask Image: Object, plant, and issue identification from photos
• Ask Audio: Private audio transcription for lectures and notes

Google demonstrated "Mobile Actions" that allow users to control core phone functions like WiFi and maps through voice or text commands, all processed locally.

Real-world applications

Google highlighted two specific use cases:

Retail: Converting recipe ideas into localized in-store shopping maps completely offline, allowing customers to navigate stores without internet connectivity.

Automotive: Providing mechanics with immediate visual diagnostics from photos of faulty parts, enabling on-the-spot troubleshooting.

What this means

This release represents Google's push to run increasingly capable AI models entirely on-device, addressing privacy concerns and enabling functionality in areas with poor connectivity. By optimizing specifically for the Tensor G5's TPU architecture, Google is differentiating its Pixel hardware through exclusive AI capabilities that competitors cannot easily replicate. The retail and automotive examples suggest Google is targeting enterprise deployments where offline operation is critical, not just consumer use cases. However, without published benchmarks or comparisons to cloud-based alternatives, the actual performance trade-offs of this on-device approach remain unclear.