A pretrained model hub that refuses to pick a backend

What it does KerasHub is a library of pretrained architectures—ResNet, BERT, GPT-2, Llama, Mistral, Whisper, and others—implemented as ordinary Keras 3 layers and models. It covers vision, text, and audio tasks like classification and generation, and hosts the corresponding weights on Kaggle Models. If you know your way around a keras.Model, the API should feel like an extension of what you already use.

The interesting bit The real trick is framework polygamy: a single model definition targets JAX, TensorFlow, or PyTorch depending on the KERAS_BACKEND environment variable, so you can train on any of the three without rewriting architecture code. The library also bundles PEFT and model-parallel fine-tuning for when you outgrow a single GPU.

Key highlights

• One model definition works across JAX, TensorFlow, and PyTorch
• Pretrained weights for vision (ResNet), NLP (BERT, Llama, Mistral), and audio (Whisper) tasks
• Built-in PEFT and model/data parallel training for fine-tuning at scale
• Standard Keras Layer and Model objects that slot into existing Keras workflows
• Weights are hosted on Kaggle Models

Caveats

• APIs are pre-release (0.y.z) and explicitly unstable; the README warns compatibility may break at any time.
• Installation currently pulls in TensorFlow because preprocessing relies on tf.data, even if you plan to train on JAX or PyTorch.
• Many underlying models (Llama, Mistral, Whisper, etc.) are provided by third parties under separate licenses.

Verdict Worth a look if you are invested in Keras 3 and want pretrained models without committing to a single backend framework. If you are already comfortable in the Hugging Face ecosystem and do not need Keras abstractions, this likely will not pull you away.