Google released Gemma 4 E2B for TPU, a version of its open model built to run natively on the Tensor Processing Unit inside the Pixel 10. The model handles offline chat, image identification, and audio transcription, and it can also execute phone functions like Wi-Fi and maps on voice or text command. That last part is the news: Gemma 4 is no longer just answering questions, it is operating the device.
AI Insiders covered Gemma 4 in June through the lens of quantization-aware training, the technique that shrank the model to roughly a gigabyte so it could run on a phone at all. Today’s release answers a different question. Now that the model fits, what does Google let it touch? The answer is core phone controls, not just a chat window.
According to 9to5Google, which broke the announcement, Google unveiled the TPU variant at I/O Connect India, a satellite event following a similar gathering in Berlin the week prior. Gemma 4 was first introduced in April as the base model for the upcoming Gemini Nano 4, and this release is the version tuned specifically for the Tensor G5 chip. It ships on the Pixel 10, 10 Pro, 10 Pro XL, and 10 Pro Fold.
Three capabilities define the release. AI Chat supports extended offline conversations, including at cruising altitude with no signal. Ask Image identifies objects, plants, or mechanical issues from a photo with no internet connection required. Ask Audio transcribes lectures and notes entirely on-device, so the recording never leaves the phone.
The feature Google is pushing hardest is Mobile Actions, demoed as letting a user “command core phone functions, like WiFi or maps, using just private voice or text,” per 9to5Google. That framing matters. Google is positioning on-device inference not as a privacy fallback for when the network drops, but as the trust layer for letting a model take actions on your behalf. A cloud assistant that controls your phone settings has to send that request somewhere first. A model running on the Tensor G5 does not.
Google’s own materials describe the model as “state-of-the-art, powerful, yet remarkably lightweight.” The announcement includes no independent benchmark results comparing Gemma 4 E2B for TPU against on-device rivals from Apple or Qualcomm-based Android competitors. Google also has not said whether Mobile Actions extends beyond Wi-Fi and maps to functions like payments or messaging, a distinction that determines whether this is a demo feature or a real replacement for tapping through settings menus.
Google pointed to retail and automotive use cases: converting a recipe into an offline in-store shopping map, or giving a mechanic a visual diagnosis from a photo of a broken part. Neither example includes adoption data, since both are demos rather than shipped products.
For teams building on-device agent features, the signal here is not the model size, it is the permission model. Google shipped a phone-control interface gated entirely to local inference before extending it to cloud-routed assistants, and any product betting on network-dependent phone control should study why.
9to5Google (Abner Li) first reported the Gemma 4 E2B for TPU release on July 14, 2026.
