LiteRT is Google's on-device runtime for machine-learning and generative-AI models, continuing TensorFlow Lite. A model authored in PyTorch, TensorFlow or JAX is converted to a .tflite file, or to .litertlm for language models, optionally quantized, and then executed on the device's CPU through XNNPACK, its GPU through the ML Drift backend, or a supported NPU.
The V2 API centres on a Compiled Model API that selects an accelerator automatically instead of requiring explicit delegates, and adds asynchronous execution, NPU distribution and I/O buffer handling. NPUs from several chipset vendors are reached through one consistent API. Alongside the core runtime, LiteRT-LM runs quantized language and diffusion models on device, LiteRT.js runs models in the browser over WebGPU and WebAssembly, a tensor-centric C++ Tensor API exposes graph and memory control, and a command-line toolkit is documented for use from coding agents.
TensorFlow Lite packages are in maintenance mode, receiving only critical security and stability updates, with active development moved to LiteRT; a migration guide covers the upgrade. Releases target a six-to-eight-week cadence with nightly builds in between. Converted models for the Gemma family, automatic speech recognition and image classification are published to a community organisation on Hugging Face.
| Platform | CPU | GPU | NPU and accelerators |
|---|---|---|---|
| Android | yes | OpenCL, OpenGL | Broadcom, Google Tensor, Intel, MediaTek, Qualcomm |
| iOS | yes | Metal | — |
| Linux | yes | WebGPU | Broadcom, Intel |
| macOS | yes | WebGPU, Metal | — |
| Windows | yes | WebGPU | Intel |
| Web | yes | WebGPU | — |
| IoT | yes | WebGPU | — |
Features
- Compiled Model API: automatic accelerator selection, asynchronous execution and efficient I/O buffer handling, without explicit delegates
- Unified NPU access: one API across NPUs from several chipset vendors
- GPU acceleration: the ML Drift backend with buffer interoperability across GPU buffer types
- Conversion pipeline: PyTorch, TensorFlow and JAX models convert to .tflite, or .litertlm for language models, with a separate quantizer
- On-device generative AI: LiteRT-LM runs quantized language and diffusion models across edge platforms from Python, C++, Kotlin, Swift and JS
- Web inference: LiteRT.js runs client-side models over WebGPU and WebAssembly
- Tensor API: a lightweight tensor-centric C++ library for graph authoring and memory control
- Command-line toolkit: a CLI documented for use inside coding-agent workflows
- Published models: Gemma, speech recognition and image classification variants distributed through a community organisation on Hugging Face
- Release cadence: nightly builds with stable releases targeted every six to eight weeks
