Google announces Gemma 4 E2B for TPU on Pixel 10, the model for generative AI without internet
Google announces Gemma 4 E2B for TPU on Pixel 10, the model for generative AI without internet
Google has introduced Gemma 4 E2B for TPU, a new variant of its open language model designed to run natively on the Tensor Processing Unit (TPU) of the Tensor G5 chip, which is onboard the Pixel 10. The announcement was made during the I/O Connect India event, just days before the Berlin stop of the same series of satellite events organized by Mountain View.
The Gemma family is Google's line of open models built to operate directly on the device, without relying on cloud servers. The fourth generation was introduced in April, when Google had already indicated it as the foundation for the future Gemini Nano 4, the model that will handle the most advanced AI functions of the upcoming smartphones in the range.
The E2B variant announced today leverages the TPU integrated in the Tensor G5. Google describes it as a model "state-of-the-art, powerful but extremely lightweight."
The support covers Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, and Pixel 10 Pro Fold.
Gemma 4 E2B for TPU is the LLM that runs natively on Pixel 10
The package of multimodal features enabled by Gemma 4 E2B includes three main tools:
- AI Chat allows in-depth conversations with the assistant even without a connection, a detail that Google highlights by mentioning use at 30,000 feet, i.e., in an airplane.
- Ask Image enables framing an object, a plant, or a technical problem and obtaining identification with 0% of data sent to the network.
- Ask Audio manages the transcription of lectures and voice notes entirely on the device, ensuring that the audio never leaves the phone.
Among the demonstrations is also Mobile Actions, the feature that allows controlling key phone functions, such as WiFi or maps, via voice or text managed locally and not routed to the cloud.
Google has indicated two concrete application scenarios. In retail, the model converts the idea of a recipe into a precise and localized map of the store's shelves, all without a connection. In the automotive field, a mechanic can obtain immediate visual diagnosis simply by photographing a faulty component.
It remains to be seen how well these functions hold up in real-world tests outside the demo scenarios: the leap from a lightweight model like E2B to a truly reliable user experience in offline conditions, with uncommon objects or noisy audio, is the true unknown. For now, Google has shown the direction; confirmation in the field will come with the daily use of the Pixel 10.