Hugging FaceとKerasHubの新統合を発表 Announcing New Hugging Face and KerasHub integration

Hugging Face and Google's Keras team have announced a native integration between KerasHub and the Hugging Face Hub, allowing Keras practitioners to load, fine-tune and re-share Transformers-format checkpoints hosted on the Hub without leaving their familiar Keras workflow.

The entry point is KerasHub's `from_preset()` method. By passing a model identifier prefixed with `hf://`, users can pull weights for popular architectures such as Llama 3, Gemma, Mistral, Qwen and PaliGemma directly from the Hub. Under the hood the checkpoint is mapped onto KerasHub's own layer implementations, so once loaded the model behaves like any other Keras model: it can be trained with `fit()`, quantized, evaluated, or used for text generation via `generate()`.

A defining feature of the integration is its alignment with Keras 3's multi-backend design. Because KerasHub models run on JAX, TensorFlow or PyTorch interchangeably, the same checkpoint can be fine-tuned on TPUs through JAX, served inside an existing PyTorch pipeline, or deployed in a TensorFlow Serving stack — all without changes to user code. After fine-tuning, models can be exported with `save_to_preset()` and pushed back to the Hub via `keras_hub.upload_preset()`, where they remain accessible to both Keras and Transformers users.

The move reflects a broader trend toward framework-agnostic interoperability. The Hugging Face Hub has become the de facto distribution layer for open-weight models, hosting hundreds of thousands of checkpoints originally produced by PyTorch-centric pipelines. Keras, meanwhile, has reinvented itself with Keras 3 by reintroducing first-class JAX and PyTorch support after years of being tied to TensorFlow, and it underpins Google's own Gemma release infrastructure. Bridging the two ecosystems lowers the friction of moving between them and arguably strengthens both: Keras gains immediate access to a much larger model catalog, while Hub authors reach an audience that favors JAX and TPU-based training.

For researchers, the practical upshot is greater freedom to pick a backend based on hardware availability or experiment requirements rather than checkpoint format. For applied teams, it likely means easier reuse of community-trained weights inside production stacks that may already standardize on Keras for its higher-level APIs. It is also worth noting that similar cross-framework efforts — such as safetensors adoption, the GGUF format in the llama.cpp world, and ONNX conversions — have shown that interoperability tends to accelerate model adoption. KerasHub's native Hub support fits squarely within that pattern, and could plausibly nudge more JAX-trained artifacts back into the broader open-source ecosystem.