Install hf-mount with Homebrew

Project links

• Homepage
• Upstream docs
• Package manager page

Commands and aliases

• hf-mount
• hf-mount-fuse
• hf-mount-nfs

Project history

Hugging Face introduced hf-mount alongside Storage Buckets as a way to attach buckets, models, and datasets from the Hub as local filesystems. The project's README describes a FUSE or NFS filesystem that fetches file bytes lazily on first read, so model and dataset consumers can browse or load remote content without cloning or downloading a full repository first.

The project sits on top of Hugging Face's newer storage direction: Buckets are S3-like mutable object storage powered by the Xet storage backend, while traditional model and dataset repositories remain Git-based and read-only when mounted. hf-mount turns both forms into a local path abstraction for training, inference, evaluation, browsing, and agent scratch-storage workflows.

Adoption history

Hugging Face's changelog announcement framed hf-mount as a way to attach remote storage much larger than local disk and explicitly called out agentic storage. The Hub access-patterns documentation then made hf-mount one of the named access methods next to volume mounts, `hf://` fsspec paths, CLI sync, and the S3 API.

Homebrew packaging gives macOS and Linux users a direct `brew install hf-mount` path. The README notes that the Homebrew package installs the NFS backend on macOS, while FUSE builds require platform support such as macFUSE or fuse3.

How it is used

The common flow is `hf-mount start bucket owner/name /mount/path` for a read-write bucket or `hf-mount start repo datasets/name /mount/path` for a read-only Hub repository. Commands can take credentials from `HF_TOKEN`, use NFS by default, or use FUSE where available.

After mounting, ordinary tools such as shells, Python libraries, model loaders, `find`, `cat`, pandas, DuckDB, training scripts, and inference servers can read paths under the mount. The README also documents cache controls, read-only mode, token-file rotation, overlay mode, and operational limits such as stale metadata windows and no cross-node locking.

Why package nerds care

hf-mount is interesting packaging-wise because it turns cloud ML storage into a traditional Unix mount workflow. The package bundles daemon and backend executables, bridges Rust filesystem code with Hugging Face storage APIs, and exposes a very package-manager-friendly surface: install a binary, run a command, get a path.

It also shows the Hugging Face CLI ecosystem expanding beyond API wrappers into host-level tools. The package matters when model artifacts are too large for laptop disks or ephemeral CI runners but the consuming program still insists on POSIX-looking files.

Timeline

• 2026-03-24: Hugging Face announced hf-mount in the Hub changelog.
• 2026: Hugging Face documented hf-mount under Storage Buckets access patterns.
• 2026: Homebrew packaged `hf-mount` for macOS and Linux installation.
• Storage Buckets era: hf-mount became one of several Hub access paths alongside `hf://` fsspec, volume mounts, CLI sync, and S3-compatible APIs.

Related projects

• hf-mount is related to Hugging Face Storage Buckets, the Hugging Face Hub, `huggingface_hub` filesystem and URI support, Xet-backed storage, FUSE, NFS, macFUSE, and ML tooling that accepts local filesystem paths for models and datasets.

Sources

• https://formulae.brew.sh/formula/hf-mount
• https://github.com/huggingface/hf-mount
• https://huggingface.co/changelog/hf-mount
• https://huggingface.co/docs/hub/storage-buckets
• https://huggingface.co/docs/hub/storage-buckets-access
• https://raw.githubusercontent.com/huggingface/hf-mount/main/README.md