Hugging Face Hub、組織向けStorage Bucketsを導入 Introducing Storage Buckets on the Hugging Face Hub

Hugging Face has rolled out Storage Buckets, a new capability that lets organizations on the Hub keep their repository data inside cloud storage they own and operate, such as AWS S3, instead of on Hugging Face's managed infrastructure. The feature is aimed squarely at enterprise users who want the collaboration experience of the Hub without giving up control over where their bytes actually live.

The model is straightforward. An organization connects its own bucket to the Hub and designates it as the backing store for selected repositories. Model weights, datasets, and other large artifacts are written into that bucket, while the Hub continues to provide the familiar surface area: repository browsing, versioning, the Datasets Viewer, Inference integrations, and the Python and JS client libraries. From a developer's perspective, workflows like push_to_hub or load_dataset largely stay the same; what changes is the underlying location of the data.

The motivation is the increasingly enterprise-grade nature of modern ML work. Frontier-scale models and multi-terabyte datasets make data residency, regional compliance (GDPR, sector-specific regulations), egress cost, and internal security review a constant friction point. By letting customers keep artifacts within their existing cloud account, Hugging Face reduces that friction and likely makes it easier for regulated organizations — finance, healthcare, public sector — to standardize on the Hub as their internal model registry.

This also fits a broader pattern in the platform's evolution. Hugging Face has been steadily investing in its storage layer, including the Xet-based content-addressed backend designed to handle very large files more efficiently than plain Git LFS. Storage Buckets can be seen as a complementary move: rather than only optimizing the managed backend, it abstracts the storage layer so that the same Hub workflows can run on top of customer-controlled infrastructure.

The approach is not unique to Hugging Face. MLOps platforms such as Weights & Biases, MLflow-based offerings, and Databricks have long supported variants of bring-your-own-bucket for artifact storage, and major model providers increasingly let customers route data through their own cloud accounts. Hugging Face adopting this pattern suggests the company is continuing to position itself less as a public model gallery and more as a serious enterprise platform for model and dataset lifecycle management — though the competitive dynamics with hyperscaler-native registries will be worth watching.

A few practical caveats apply. Based on the announcement, Storage Buckets appears to be targeted primarily at Enterprise Hub organizations, and details around supported providers, IAM and permission design, encryption, and how billing interacts with customer-side storage costs should be verified in the official documentation. Teams considering migration should also think about how existing repositories, LFS objects, and Xet-backed files map onto the new arrangement. Still, for organizations that previously could not push proprietary data to a SaaS hub at all, this likely represents a meaningful unlock: the Hub's collaboration features without surrendering custody of the underlying data.