Google Tensor ML SDK 测试版发布
阅读原文· developers.googleblog.comGoogle 把 TPU 塞进了 Pixel 10,现在开发者能直接在手机上跑 Gemma 3 了,这是移动端 AI 从「能用」到「好用」的关键一步,做 app 的值得关注。
Google Tensor ML SDK 进入测试版,支持开发者直接在 Pixel 10 设备的 TPU 上构建和部署高性能机器学习模型。该 SDK 集成边缘部署框架 LiteRT,提供统一工作流,可高效转换、编译并运行 PyTorch 或 TFLite 模型,并具备稳定回退机制。此外,新推出的模型库包含超过 100 个经典及生成式 AI 模型(如 Gemma 3),支持低延迟、注重隐私的语音识别、计算机视觉与文本生成等功能。
Google Tensor SDK Beta with LiteRT
Google Tensor ML SDK empowers you to build on-device machine learning capabilities starting from Google Pixel 10 family of devices1, while leveraging Pixel's custom-designed Google Tensor System-on-Chip (SoC) with its dedicated Tensor Processing Unit (TPU) inference accelerator. Today, the Tensor ML SDK is graduating from its Experimental Access Program (EAP) to Beta, enabling developers to build and deploy their AI experiences seamlessly on Google Tensor’s TPU. Tensor’s TPU unlocks interactive, realtime and private on-device AI experiences such as Pixel’s Pro Zoom2, Add Me, Voice Translate3 and Call Notes4 to name a few.
Link to Youtube Video (visible only when JS is disabled)
The Beta launch unlocks two major advantages for developers:
- A Unified Developer Workflow with LiteRT
- A Model Garden to run 100+ models optimally on Tensor’s TPU
A Unified Developer Workflow with LiteRT
LiteRT is Google's on-device framework for high-performance machine learning (ML) deployment on edge platforms. It abstracts away low-level, vendor-specific SDKs including compilers and runtimes and exposes them through a unified, streamlined developer-facing API. Tensor ML SDK is now integrated with LiteRT offering a seamless developer workflow to convert, compile, deploy and run ML and Generative AI models on Google Pixel via Tensor’s TPU.
- Compilation: Convert and compile your PyTorch or TFLite models into optimized binaries ready to leverage the Tensor TPU using LiteRT Torch.
- Deployment: Use Play Feature Delivery to distribute and install the compatible runtime and compiler libraries that connect to the on-device TPU drivers. Use AI Packs (part of Play for On-device AI) to bundle and deliver the compiled model files within your application.
- Run Inference: Leverage LiteRT Runtime to run your model on the TPU with just a few lines of code. It also allows you to enable robust fallback mechanisms by specifying CPU or GPU as secondary options, and automatically uses them depending on TPU availability.
For a full, detailed guide, including colab and sample apps, visit the LiteRT NPU documentation.
Model Garden of 100+ models on Tensor
The Tensor SDK Beta enables deployment of a wide array of models—including computer vision and speech recognition—straight to Pixel devices through a large model garden. The model garden provides over 100+ Classic ML models and Generative AI Models (Gemma 3 1B), alongside a library of precompiled models available to download directly from the LiteRT Hugging Face community.
Here is a look at what developers can build with the models available:
- Small Language models: Enable local actions for app interactions with Function Gemma. Add rich semantic features to your app with EmbeddingGemma.
- Intelligent Content creation: Build features that generate real-time text, apply smart image filters, and execute advanced computational photography effects—like portrait blurring
- Vision & Understanding: Implement object detection, depth mapping, body tracking, multimodal image to text understanding to power camera applications that recognize and react to the user’s environment
- Audio & Accessibility: Run end to end speech recognition to deliver secure, low latency audio transcription, voice-controlled accessibility tools, and translation on the edge
Here are a couple of demos for inspiration:
Get Started with Pixel Today
We invite the developer community to build a new class of intelligent and responsive AI applications starting with the Pixel 10 family. Explore the Tensor SDK, experiment with running your models on the TPU, and share your feedback with the LiteRT community on Github.
If you're joining us at I/O—either in person or virtually—be sure to check out our dedicated codelabs and sample app. You'll get hands-on experience using the Tensor SDK, compiling models, and deploying them to the family of Pixel 10 devices via TPU.
- Documentation of Google Tensor SDK with LiteRT
- Browse the 100+ Models in the Model Garden
- Download Precompiled Models on LiteRT Hugging Face community
- Start the Google I/O Tensor SDK Codelab
- Review the Tensor ML SDK License and Distribution terms
Devices Supported: Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, and Pixel 10 Pro Fold.
We can’t wait to see what you build on Pixel!
Acknowledgements
This project was made possible through the collaboration of several teams. We thank them for their significant contributions:
Tensor Team: Himangshu Roy, Chirag Gupta, Rishubh Khurana, Rachit Agrawal, Priya Patel, Prakul Sawhney, Abby Chung, Malini P V, Jui Pradhan, Naina Singla, Debapriya Maji, Aditya Srivastava, Abhishek Jatram, Vibhu Agrawal, Rachana Jayaram, Lokesh Vutla, Abhishek Singh, Annie Fu, Chen-Hao Liao, Chanchal Raj, Ty Werbicki, Sriram Kashyap M S, Shubham Saini, Thiru Ramasamy, Jayanthan K, Payal Agarwal, Pranjal Srivastava, Yathish Reddy M, Akhilesh Ravi, Harold Yang, Yi Yo, Priyanka Mittal, Ishaan Agrawal, Vivek Kumar, Minje Park, YoonKyung Kim, Eunji Heo, Mehran Nekuii, John Joseph, Nett Phasukavanich, Jess Tsopanis, Jeff Setter, Ganesh Rao, Neena Maldikar, Yohsin Fang, Estelle Liao, Leo Wu.
LiteRT Team: Lu Wang, Weiyi Wang, Jingjiang Li, Gerardo Carranza, Terry (Woncheol) Heo, Andrew Zhang, Chenchen Tang, Shuangfeng Li, Changming Sun, Somdatta Banerjee, Na Li, Yu-hui Chen, Tenghui Zhu, Alice Zheng, Chintan Parikh, Sachin Kotwani, Cormac Brick, Matthias Grundmann, Salil Tambe, Yishuang Pang.
Licensing
https://ai.google.dev/edge/litert/next/tensor_ml_terms
1 - Devices supported are Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, and Pixel 10 Pro Fold.
2 - Available in select countries and languages.
3 - Available in select countries and languages. Results may vary. Check responses for accuracy.
4 - Available in the US. English Only
Explore this announcement and all Google I/O 2026 updates on io.google.