TensorFlow Lite and LiteRT for Mobile

TensorFlow Lite has been renamed LiteRT in current Google AI Edge documentation, while existing apps and model artifacts may still use TensorFlow Lite terminology. The naming transition does not mean every application must rewrite its code immediately; follow migration guidance for the specific runtime and API being used. Older tutorials may refer to Interpreter APIs, while current docs also describe newer compiled-model paths. A mobile workflow begins with a trained model and a conversion step that produces a supported on-device representation. Conversion can include optimization or quantization. The result must be checked for compatible operators, input shapes, tensor types, and metadata. If conversion reports unsupported operations, a model may need export changes, custom operations, or a different runtime path. At inference time, LiteRT can execute on CPU and may use delegates for supported GPU or NPU hardware. A delegate accelerates operations it supports; unsupported parts may fall back to another device or fail depending on configuration. Availability varies across device models, chipset vendors, runtime packaging, and operating-system versions. Do not assume an NPU delegate is present simply because the phone contains an NPU. Quantization reduces numeric precision and can shrink model size or improve speed, but may change accuracy. Calibration data should represent the intended input distribution for supported post-training quantization methods. Test both resource use and task metrics after conversion. Image color order, normalization, tokenizer settings, and output decoding need to match training. On-device deployment also involves app lifecycle, memory pressure, battery use, installation size, privacy, and version support. Benchmark on physical target devices, including startup and sustained runs. Keep model version and conversion configuration with the app release. LiteRT is a runtime, not a guarantee that a model will be accurate or compatible with every phone.