keras-hub

Why we included this project

Teams already building with Keras get a shortcut here: ready-made implementations of popular text, vision, and audio models as ordinary Keras layers, alongside pretrained checkpoints hosted on Kaggle Models. You can load something like a BERT classifier or a ResNet image model instead of re-implementing the architecture yourself. The stronger draw is the single model definition, which runs unchanged on TensorFlow, JAX, or PyTorch, so a team standardizing on Keras 3 can still move across frameworks or accelerators without rewriting its pipeline. Because the pieces are standard layers and models, fine-tuning a pretrained model for a new task follows the Keras workflow you already know, and built-in PEFT support covers parameter-efficient training on a single GPU.