GOOGLE RENAMES TENSORFLOW LITE TO LITERT AND INTRODUCES A COMPILED INFERENCE PATH
Google renamed TensorFlow Lite to LiteRT, added a new CompiledModel API, and put the old TFLite packages in maintenance mode. LiteRT keeps the .tflite format a...
Google renamed TensorFlow Lite to LiteRT, added a new CompiledModel API, and put the old TFLite packages in maintenance mode.
LiteRT keeps the .tflite format and Interpreter API but adds a new CompiledModel path that selects GPU/NPU via options instead of manual delegates. See the rename details and version guidance in this clear rundown: LiteRT vs TensorFlow Lite.
There’s also a PyTorch converter litert-torch that goes from torch.export straight to .tflite, and a separate LiteRT-LM runtime replacing the MediaPipe LLM API. Existing TFLite apps keep working, but future updates land in LiteRT.
TensorFlow Lite is effectively frozen; ongoing on-device work will move to LiteRT.
CompiledModel and litert-torch simplify acceleration and model conversion, reducing glue code and toolchain hops.
-
terminal
A/B latency and battery on target devices: Interpreter vs CompiledModel with Accelerator.GPU/NPU using the same .tflite.
-
terminal
Export a model via torch.export -> litert-torch vs your current ONNX/TFLite path; compare accuracy, size, startup, and build time.
Legacy codebase integration strategies...
- 01.
Migrate incrementally: swap Gradle and pip packages first, keep .tflite and Interpreter, then move hot paths to CompiledModel.
- 02.
Map existing GPU/NNAPI delegates to CompiledModel options; verify minSdk and check gaps (support/metadata 2.x, Task Library still under TFLite).
Fresh architecture paradigms...
- 01.
Start on LiteRT 2.2.x with CompiledModel and design a capability switch for NPU/GPU per device.
- 02.
Prefer PyTorch torch.export -> litert-torch to avoid ONNX/TF graphs and simplify CI.
Get daily GOOGLE + SDLC updates.
- Practical tactics you can ship tomorrow
- Tooling, workflows, and architecture notes
- One short email each weekday