借助 Google Tensor 与 Pixel 开启设备端 AI 新纪元

印度 Google I/O 圆满结束！在与 Google Pixel 团队的特别联合展示中，我们正式揭晓了 Google Tensor 的定制片上系统（SoC）与先进的张量处理单元（TPU）如何驱动全新 Google Pixel 10 系列设备[^1] 实现强大且 100% 私密的下一代设备端 AI。

以下是本次印度 Google I/O 具有开创性的展示内容概要：

面向 TPU 的 Gemma 4 E2B 正式发布

我们首次推出 面向 TPU 的 Gemma 4 E2B—— 一款先进、强大但极为轻量的模型，专为在 Pixel TPU 上原生运行而设计。Gemma 4 可确保您的数据永不离开设备，为边缘 AI 应用提供绝对的离线隐私保护。

Agent 技能与移动操作

我们演示了设备上的 TPU 如何在完全离线状态下充当个人旅行规划师、食谱指南或家庭自动化管家。此外，借助在 TPU 上原生运行的 FunctionGemma，用户可以通过移动操作执行核心手机功能——例如控制 Wi‑Fi 或地图——只需使用私密的语音或文本。

丰富的离线多模态交互

我们展示了无需联网即可无缝进行文本、图像与音频的推理：

AI 聊天： 即时、深度离线对话（即使在三万英尺高空）。
图像问答： 在零网络连接下拍摄并识别物体、植物或问题。
音频问答： 对讲座和笔记进行 100% 私密的设备端音频转录。

真实的边缘应用案例

我们重点介绍了该技术如何立刻推动行业变革：

零售： 将食谱创意转化为完全离线的精确本地化店内购物地图。
汽车： 让修理工通过拍摄故障零件照片即可现场进行视觉诊断。

使用 Tensor SDK 进行构建

注册 Tensor SDK，您即可获得统一开发者工作流和 100 余个经典机器学习模型，以及构建安全边缘 AI 所需的最新 SLM 模型。请参考我们的 开源 Edge TPU 应用代码库，即日起开始在 Pixel 上构建下一代 AI。

很抱歉，您的浏览器不支持播放此视频

开源 Edge TPU 应用

重要链接

致谢

本次 I/O 活动由多个团队协作促成。我们感谢他们的重要贡献：

参与印度 Google I/O Connect 的 Tensor 团队： Himangshu Roy、Chirag Gupta、Rishubh Khurana、Shubham Saini、Priya Patel、Prakul Sawhney、Malini P V、Chanchal Raj、Abhishek Jatram、Akhilesh Ravi、Jui Pradhan、Naina Singla、Jayanthan K、Sarvesh Das、Alice Zheng、Yishuang Pang、Na Li、Lu Wang、Vibhu Agrawal、Debapriya Maji、Aditya Srivastava

Pixel 业务团队： Lokesh Madala、Amit Roykaran、Mitul Shah、Keya Mukherji、Sneha Budhia

许可

https://ai.google.dev/edge/litert/next/tensor_ml_terms

[^1]: 支持的设备：Pixel 10、Pixel 10 Pro、Pixel 10 Pro XL、Pixel 10 Pro Fold

That's a wrap on Google I/O India! In a special joint showcase with the Google Pixel team, we officially unveiled how Google Tensor’s custom System-on-Chip (SoC) and advanced Tensor Processing Unit (TPU) are driving the next generation of powerful, 100% private on-device AI for from Google Pixel 10 family of devices1

Here is a summary of the groundbreaking Google India I/O Showcases:

Introducing Gemma 4 E2B for TPU

We debuted Gemma 4 E2B for TPU—a state-of-the-art, powerful, yet remarkably lightweight model designed to run natively on the Pixel TPU. Gemma 4 ensures your data never leaves the device, guaranteeing absolute offline privacy for edge AI applications.

Agent Skills & Mobile Actions

We demonstrated how your on-device TPU can act as a personal travel planner, recipe guide, or home automator completely offline. Additionally, using FunctionGemma running natively on the TPU, users can execute Mobile Actions to command core phone functions—like WiFi or maps—using just private voice or text.

Rich Offline Multimodal Interactions

We showcased seamless, internet-free reasoning across text, images, and audio:

AI Chat: Instant, deep offline conversations (even at 30,000 feet).
Ask Image: Snap and identify objects, plants, or issues with 0% internet.
Ask Audio: 100% private, on-device audio transcription for lectures and notes.

Real-World Edge Applications

We highlighted how this technology transforms industries right now:

Retail: Converting recipe ideas into precise, localized in-store shopping maps completely offline.
Automobile: Giving mechanics on-the-spot visual diagnostics from photos of faulty parts.

Build with the Tensor SDK

Sign up for the Tensor SDK, offering the unified developer workflow and 100+ classical ML models in addition to the latest SLM models you need to build secure, edge-based AI. Refer to our Open Source Edge TPU Application code base and start building next-gen AI on Pixel today.

Sorry, your browser doesn't support playback for this video

Open Source Edge TPU Application

Important Links

Acknowledgements

This I/O Event was made possible through the collaboration of several teams. We thank them for their significant contributions:

Tensor Team for Google I/O Connect India Event: Himangshu Roy, Chirag Gupta, Rishubh Khurana, Shubham Saini, Priya Patel, Prakul Sawhney, Malini P V, Chanchal Raj, Abhishek Jatram, Akhilesh Ravi, Jui Pradhan, Naina Singla, Jayanthan K, Sarvesh Das, Alice Zheng, Yishuang Pang, Na Li, Lu wang, Vibhu Agrawal, Debapriya Maji, Aditya Srivastava

Pixel Business Team: Lokesh Madala, Amit Roykaran, Mitul Shah, Keya Mukherji, Sneha Budhia

Licensing