Google Tensor SDK Beta with LiteRT

The Google Tensor ML SDK lets you build on-device machine learning capabilities from the Google Pixel 10 family of devices.¹while leveraging Pixel’s custom-designed Google Tensor system-on-chip (SoC) with dedicated Tensor Processing Unit (TPU) inference accelerators. The Tensor ML SDK has now graduated from the Experimental Access Program (EAP) and into beta, allowing developers to seamlessly build and deploy AI experiences on Google Tensor’s TPU. Tensor’s TPU enables interactive, real-time, private, on-device AI experiences like Pro Zoom on Pixel²please add voice translation³ and call notes⁴ To name a few examples.

Our product management team details Tensor SDK and what it means for the future of Pixel development.

Beta releases offer two major benefits for developers:

Integrated developer workflow with LiteRT
Model Garden for optimally running 100+ models on Tensor’s TPU

Integrated developer workflow with LiteRT

LiteRT is Google’s on-device framework for deploying high-performance machine learning (ML) on edge platforms. Abstract low-level vendor-specific SDKs, including compilers and runtimes, and expose them through a unified, streamlined developer API. The Tensor ML SDK integrates with LiteRT to provide a seamless developer workflow to transform, compile, deploy, and run ML and generative AI models on Google Pixel via Tensor’s TPU.

edit: Convert and compile your PyTorch or TFLite model into an optimized binary to take advantage of Tensor TPUs using LiteRT Torch.
introduction: Use Play Feature Delivery to distribute and install compatible runtime and compiler libraries that connect to on-device TPU drivers. Use AI Pack (part of Play for On-device AI) to bundle and deliver compiled model files within your application.
perform inference. Run your model on the TPU with just a few lines of code using the LiteRT runtime. You can also enable a robust fallback mechanism by specifying CPU or GPU as a secondary option and automatically use them depending on TPU availability.

For a complete detailed guide, including colab apps and sample apps, see the LiteRT NPU documentation.

Model garden for over 100 models on Tensor

Tensor SDK Beta allows you to deploy a wide range of models, including computer vision and speech recognition, directly to Pixel devices through a large model garden. What the model garden provides: 100+ classic ML models Contains Generative AI Models (Gemma 3 1B) and a library of precompiled models that can be downloaded directly from the LiteRT Hugging Face community.

Here’s what developers can build with the available models.

small language model: Enable local actions for app interaction with Function Gemma. Add rich semantic functionality to your app using EmbeddingGemma.
Intelligent content creation: Build the ability to generate real-time text, apply smart image filters, and perform advanced computational photography effects like portrait blur.
vision and understanding: Implement object detection, depth mapping, body tracking, and multimodal image-to-text understanding to enhance camera applications that recognize and react to the user’s environment.
Audio and accessibility: Perform end-to-end speech recognition for secure, low-latency speech transcription, voice-controlled accessibility tools, and translation at the edge.

Here are some demos for inspiration.

Sorry, your browser does not support playing this video

Image Segmentation LiteRT Sample App (left) Function Gemma with LiteRT LM on AI Edge Gallery – Mobile Action (right)

Get started with Pixel today

We’re calling on 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 by running your models on TPU, and share your feedback with the LiteRT community on Github.

If you’re attending I/O in person or virtually, be sure to check out our dedicated codelabs and sample apps. Gain hands-on experience using the Tensor SDK, compiling a model, and deploying to the Pixel 10 device family via TPU.

Supported devices: Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, Pixel 10 Pro Fold.

We can’t wait to see what you build with Pixel.

Acknowledgment

^{This project was made possible by the collaboration of multiple teams. We would like to thank them for their great contributions.}

^{tensor team}^{: Himanshu Roy, Chirag Gupta, Rishubh Khurana, Rachit Agrawal, Priya Patel, Prakul Sawhney, Abhi Chan, Malini PV, Jui Pradhan, Naina Singla, Devapriya Majhi, Aditya. Srivastava, Abhishek Jatram, Vibhu Agrawal, Rachana Jayaram, Lokesh Vutla, Abhishek Singh, Annie Hu, Chenghao Liao, Chanchal Raj, Tai Werbicki, Shriram Kashyap. MS, Shubham Saini, Thiru Ramasamy, Jayanthan K, Payal Agarwal, Pranjal Srivastava, Yatish Reddy M, Akhilesh Ravi, Harold Yang, Yi Yo, Priyanka Mittal, Ishaan Agrawal. , Vivek Kumar, Min-Jae Park, Yoon-Kyung Kim, Eunji Ho, Mehran Nekui, John Joseph, Net Faskavanich, Jess Tsopanis, Chenhao Liao, Jeff Setter, Ganesh Rao, and Nina Mardikar.}

^{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 Tanbe.}