.TFLITE

Introduction

A .tflite file is a trained model prepared to run on phones, embedded boards and microcontrollers with TensorFlow Lite, which Google renamed LiteRT in 2024. The converter takes a TensorFlow, Keras or JAX model, fixes its graph, and often quantises the weights to 8-bit integers so the file is smaller and runs faster on small chips. The result holds both the network and its weights in one file that can be mapped into memory and used without parsing.

Structure

The file is a FlatBuffers buffer whose schema, schema.fbs, is part of the TensorFlow source. It starts with a 4-byte offset to the root table, then the file identifier TFL3 at bytes 4 to 7.

The Model table has a schema version, a list of operator codes such as CONV_2D and FULLY_CONNECTED, one or more subgraphs, a list of buffers and optional metadata. Each subgraph lists its tensors (with type, shape, quantisation parameters and the buffer that holds their values), its inputs and outputs, and its operators in the order they run. Large models may keep buffer data after the FlatBuffer instead of inside it.

Links