The runtime built for not-a-server
TensorFlow Lite (TFLite) runs TF models on mobile, embedded, and edge devices — phones, microcontrollers, Raspberry Pi, cameras. The full TF runtime weighs tens of MB and assumes float32 compute. TFLite is stripped down to run in kilobytes of RAM on processors without floating-point hardware.
The workflow has two phases. First, train a model with the full TF/Keras stack on a powerful machine. Then convert it to the .tflite FlatBuffer format with TFLiteConverter. The result is self-contained — graph + weights in one file — loaded by the TFLite Interpreter on device.
LiteRT migration: from TF 2.19+,
tf.lite.Interpreter is being superseded by ai_edge_litert.interpreter from the separate LiteRT repository. TFLite is becoming its own project. Existing code keeps working, but new code should target LiteRT for long-term maintenance.