Keras Hub

The pitch here is subtraction, not addition. KerasHub gives you Keras 3 implementations of well-known architectures plus pretrained checkpoints on Kaggle Models, and then gets out of the way. The quickstart is two short blocks: load resnet_50_imagenet into an ImageClassifier and predict an image, or load bert_base_en_uncased into a BertClassifier with num_classes=2 and call .fit() on the IMDb reviews dataset pulled via TensorFlow Datasets. Nothing new to learn about training loops, serialization, or callbacks — the developer guides cover those once for Keras and they apply unchanged. The multi-backend story is the part that separates it from a plain model zoo. Setting KERAS_BACKEND to "jax", "tensorflow" or "torch" changes where training and inference execute without touching your model code, which matters if you are benchmarking JAX against your existing TensorFlow setup or sharing code with a team standardized on Torch. The honest cost is stated in the installation docs: pip install keras-hub always pulls in TensorFlow for the tf.data preprocessing path, so your environment grows even if you train on JAX. That is a real dependency, not a footnote. Where it does not fit: the catalog is whatever lives on Kaggle Models, which is nothing like the breadth of Hugging Face Transformers, and the API is pre-release 0.y.z. The project says it plans Semantic Versioning and backwards-compatibility guarantees for both code and saved models, but until that lands you should pin versions and expect churn. Keras 2 users also need to migrate first. As a fast on-ramp to transfer learning inside Keras 3, it does its job with very little ceremony.