Announcing New Hugging Face and KerasHub integration

Source: HuggingFace Date: 2024-07-10 URL: https://huggingface.co/blog/keras-hub-integration

Summary

Integration announcement: HF Hub and KerasHub now share a common model save format, enabling KerasHub users to load any of the 300K+ Transformers-format models via from_preset("hf://...") syntax. Initial architecture support: Gemma 1/2, Llama 3, PaliGemma. Supports TensorFlow, JAX, and PyTorch backends from the same checkpoint. No performance benchmarks.

Implications

Thread: HF as open-source ML hub. This is an ecosystem expansion play — by making HF Hub the common distribution layer for both Transformers and KerasHub users, HF extends its reach to the Google/TensorFlow/JAX developer community that uses Keras. For JAX practitioners in particular (who previously had limited access to the Transformers model ecosystem), this opens up the full HF model library. The shared safetensors format is the enabling technical piece. Low ongoing strategic signal — primarily valuable at launch for TF/JAX users discovering the integration.