tensorflow

First released in 2015, TensorFlow introduced computational graph-based deep learning to a broad developer audience and powered production ML systems at Google scalefrom search ranking to translation to image understanding.

Its ecosystem spans TensorFlow Core for model building and training, TensorFlow Lite for on-device inference, TensorFlow.js for browser-based ML, and TensorFlow Extended (TFX) for production ML pipelines.

The framework's evolution from TF1's static graph execution to TF2's eager execution (mirroring PyTorch's imperative style) reflects the broader shift in the deep learning community toward more Pythonic, interactive model development.

Keras, integrated as TensorFlow's high-level API in TF2, provides an accessible layer for building and training models with minimal boilerplate, while still allowing researchers to drop into lower-level graph operations when architectural experimentation requires it.

TensorFlow's SavedModel format and TensorFlow Serving enable reliable production deployment of trained models.

Enterprise ML teams with established TensorFlow infrastructure, Google Cloud customers using Vertex AI and TPU acceleration, mobile developers deploying models via TensorFlow Lite, and organizations with existing TFX pipelines continue to use TensorFlow as their primary framework.

While PyTorch has become the dominant framework in research, TensorFlow maintains strong production deployment adoption, particularly at companies that built their ML systems before PyTorch's research dominance and have accumulated significant TensorFlow-specific infrastructure, tools, and organizational expertise.