Cloudsmith Hugging Face Registry

Control and visibility over every Hugging Face model in your organization

Teams pull models from Hugging Face like any public registry. Cloudsmith proxies, caches, and secures those requests - giving you one place to manage open source and proprietary AI artifacts alongside all your other packages and containers.

Control every artifact used in your AI projects. Cloudsmith's Hugging Face support gives your teams a secure, private registry for models and datasets.

Cloudsmith extends artifact management to AI/ML, giving your teams a secure, private registry for models and datasets. Developers simply point their projects at Cloudsmith as they would to Hugging Face Hub. Enterprises gain governance, security, and global delivery at scale. With proxying and caching for Hugging Face Hub, organizations can adopt external models confidently and build on them securely.
    The single source of truth for your machine learning models and datasets
    Store and manage models and datasets sourced from Hugging Face Hub alongside 30+ other artifact formats. Cloudsmith becomes the single source of truth for every artifact and asset used in your AI workflows, all governed by the same access controls, scanning policies, and audit logging.
    Use Cloudsmith just like Hugging Face Hub
    Cloudsmith is your own, private Hugging Face Hub. Fully compatible with the huggingface_hub Python library and the Hugging Face CLI, letting you push and pull models and datasets with the same commands you already use. Simply set HF_ENDPOINT to your Cloudsmith repository and your existing workflows continue without modification.
Universal format support

Hugging Face models and datasets, plus 30+ formats, one registry. Cloudsmith is the secure, centralised store for all your AI artifacts and software packages.

  • Store and version models and datasets using familiar Hugging Face tooling
  • Proxy and cache models and datasets from Hugging Face Hub via Cloudsmith upstreams
  • Manage ML models alongside containers, Helm charts, and language packages in one place

Signs you're ready to bring Hugging Face models under proper governance

If your teams are pulling models directly from Hugging Face Hub without visibility, policy enforcement, or version control, Cloudsmith gives you the governance layer your AI supply chain needs.
    Teams are pulling models directly from Hugging Face Hub
    When developers pull models directly from the public Hub, they bypass scanning, policy checks, and audit trails. Cloudsmith's upstream proxying intercepts those requests, caches the model, and enforces your enterprise policies before the artifact reaches any workflow.
    You have no visibility into which model versions are in use
    Without versioning and provenance tracking it is nearly impossible to prove which model is running in production, where it came from, or what depends on it. Cloudsmith maps every Hugging Face commit to an immutable, versioned package so you always know exactly what is deployed and can reproduce any build.
    Unsafe model file formats are entering your pipeline
    Formats such as Pickle and PyTorch checkpoints are common in Hugging Face models and carry well-known serialisation vulnerabilities. Cloudsmith's Enterprise Policy Manager lets you write OPA Rego policies to quarantine or block models containing risky file formats before they reach developers or production.
    Model licences are uncontrolled and legally risky
    Open-weight models like Llama and Gemma carry custom licence terms that restrict commercial use, redistribution, and competitive training. Cloudsmith parses model card metadata and exposes licence information to policy rules, so only models with approved licences can enter your pipeline.
    Your ML artifacts are isolated from the rest of your supply chain
    Managing models in a separate system from your containers, language packages, and binaries creates policy gaps and audit blind spots. Cloudsmith unifies Hugging Face models and datasets with all other artifact formats under one governed, observable platform with consistent access controls and audit logging.

Get started with Hugging Face on Cloudsmith

Frequently asked questions

  1. Yes. Cloudsmith is fully compatible with the huggingface_hub Python library and the Hugging Face CLI. To point your tooling at Cloudsmith, set the HF_ENDPOINT environment variable to your Cloudsmith repository URL and set HF_TOKEN to your Cloudsmith API token. From that point, all push, pull, and download commands work identically to how they work against Hugging Face Hub, with no changes to your existing workflows.

  2. Cloudsmith supports Hugging Face model repositories and dataset repositories. You can push and pull models (including weights, configs, tokenizer files, and model cards) and datasets using the standard huggingface_hub library with repo_type set to 'model' or 'dataset'. Hugging Face Spaces are not currently supported.

  3. Yes. You can configure a Cloudsmith repository as an upstream proxy for the public Hugging Face Hub. Model and dataset pull requests flow through your private Cloudsmith repository, which caches artifacts locally. This gives you faster, more reliable access for your teams and CI/CD pipelines, while allowing you to apply enterprise policies to any model before it reaches a developer or production environment.

  4. Cloudsmith maps Hugging Face repository concepts directly to its own package model. A Hugging Face repository becomes a Cloudsmith package, and each commit becomes a distinct, immutable package version identified by its commit hash. Tags such as 'main', 'v1.0', or 'latest' are also supported and resolve to the correct underlying commit automatically. This ensures every model version is traceable, reproducible, and tamper-evident.

  5. Cloudsmith's Enterprise Policy Manager (EPM) lets you write OPA Rego policies that target attributes specific to Hugging Face artifacts. Cloudsmith parses model card metadata, exposing fields such as licence type, training data provenance, and file formats to your policy rules. Common policy patterns include blocking models that contain risky serialisation formats like Pickle, restricting downloads to models from approved publishers, and enforcing licence compliance before any model enters production.

  6. Yes. Cloudsmith's Hugging Face repositories use native file storage deduplication. If the same model or dataset files are uploaded multiple times, even across different repositories within the same workspace, Cloudsmith stores only a single copy of the data. This significantly reduces storage consumption and associated costs for large model weights and datasets.

  7. Authentication uses standard Hugging Face tooling. Set the HF_TOKEN environment variable to your Cloudsmith API token before running any huggingface_hub or hf CLI commands. Cloudsmith also supports entitlement token authentication for distributing models to downstream consumers, and OIDC for CI/CD environments such as GitHub Actions and Jenkins.

  8. Yes. Cloudsmith supports 30+ formats in a single platform. You can store Hugging Face models and datasets alongside Docker container images, Python packages, npm packages, Helm charts, Maven artifacts, and more. All formats share consistent access controls, policy enforcement, vulnerability scanning, and audit logging, giving you a truly unified software supply chain.

  9. Yes. You can push models and datasets to Cloudsmith using the huggingface_hub upload_folder function or the hf upload CLI command, pointing HF_ENDPOINT at your Cloudsmith repository. For organisations sourcing models from the public Hub, configuring Cloudsmith as an upstream proxy will incrementally cache models as they are pulled, removing the need for a one-time bulk migration.

  10. Yes. Because Cloudsmith is fully compatible with the Hugging Face SDK and CLI, you can integrate it into any CI/CD pipeline that already pulls models or datasets. Set HF_ENDPOINT and HF_TOKEN in your pipeline environment and your existing model download steps will resolve through Cloudsmith automatically, giving you caching, policy enforcement, and audit logging without changing your pipeline code.