// the find
tensorflow/tensorflow
An Open Source Machine Learning Framework for Everyone
TensorFlow is Google's ML framework, now over a decade old and carrying that history. It's the safe enterprise choice for production ML pipelines where JAX or PyTorch would require justification to management. Most new research isn't happening here anymore.
The TFLite and TFX ecosystem is genuinely production-ready for mobile and pipeline work in ways PyTorch still isn't. The C API surface is stable and well-specified, which matters if you're embedding ML in a non-Python runtime. Distributed training via tf.distribute works reliably at scales where you'd otherwise be writing a lot of custom code. The XLA compiler integration means you can get real performance gains without leaving the framework.
The eager/graph duality is a decade-long design debt that never got fully resolved — tf.function tracing bugs are a time sink that PyTorch users don't face. Keras went through an architectural divorce (keras 3.x is now a separate project) and the migration story is rough. Building from source is a Bazel-flavored nightmare; the prebuilt wheels have a narrow CUDA version matrix and you'll hit a wall the moment your cluster runs a slightly different driver. New model architectures almost always debut in PyTorch first, sometimes never making it to TF at all.