Google Launches LiteRT Tensor API for Native On-Device ML Development
TorchTPU lets developers train, serve and deploy AI models on Google’s chips with minimal code changes, while Google also added tools for moving apps to its cloud.
- On Tuesday, Google announced the open-source release of TorchTPU at PyTorch Conference 2026, enabling developers to train, serve, and deploy AI models on Google's TPUs using familiar PyTorch APIs.
- PyTorch, the popular toolkit used by 85% of AI apps, was originally developed by Meta in 2016; its features were built to work specifically with Nvidia's GPUs, which became the standard hardware for AI research.
- Google worked with Meta, Hugging Face, and other developers to ensure TorchTPU's code libraries function seamlessly, and launched a dedicated TPU agent suite to help developers move apps to Google Cloud.
- Amazon released its own version of PyTorch, called TorchNeuron, on Monday, allowing developers to port AI apps to Amazon's homegrown Trainium chips and AWS with no rewriting required.
- By making proprietary chips easier to use with PyTorch, Google and Amazon seek to reduce software barriers, offering developers greater flexibility in choosing computing hardware without substantially modifying existing code.
15 Articles
15 Articles
Google Launches LiteRT Tensor API for Native On-Device ML Development
Google's new LiteRT Tensor API lets developers hand-build on-device models op by op in C++ and run the same graph on CPU, GPU or NPU, or export a .tflite file. Google still recommends converting trained models for most projects; the Tensor API targets developers who want low-level control.
The new solutions allow you to migrate PyTorch-based applications to your own processors without rewriting code.
Just yesterday we reported on Torch Neuron, a new tool that makes it easy for companies to move their AI apps to Amazon's Tranium AI chip. Now it is clear that Google has released a similar tool called Torch TPU, but with the difference that the AI apps are moved to Google's TPU chip (Tensor Processing Unit). The hope is that the new tool will help Google challenge Nvidia, which is currently the market leader in the segment. According to Techcru…
TorchNeuron (Amazon) and TorchTPU (Google), published in open source on 20 October 2026, execute on Trainium and TPU the PyTorch code written for the Nvidia GPU by changing the target device. The handwritten CUDA and Triton kernels are not supported: NKI at Amazon, Pallas and Tokamax at Google, Helion still experimental. TorchNeuron is in public beta, requires PyTorch 2.14.0 and does not accept dynamic shapes. At Deepgram, the 11-70% additional …
Coverage Details
Bias Distribution
- 67% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
