Local LLM Manager

A terminal UI for managing your local HuggingFace model cache — browse, inspect, delete, and download models/datasets/spaces without leaving the terminal.

Features

• Browse all cached models, datasets, and spaces with size visualization
• Inspect revisions, branches/tags, and individual files per model
• Delete entire repos or specific revisions (orphaned blobs cleaned up automatically)
• Download new models using hf download directly from the TUI
• Search and filter by name, type (model/dataset/space), sort by size, name, or date
• Scrollable list with adaptive layout to any terminal width

Requirements

• Bun — runtime and bundler
• HuggingFace Hub CLI for downloading models: pip install huggingface_hub

Install & Run

git clone https://github.com/afanjul/local-llm-manager
cd local-llm-manager
bun install
bun start

Build standalone binary

Requires ldid for ad-hoc signing of the ARM64 binary:

brew install ldid

bun run build    # → ./local-llm-manager

The binaries are self-contained — no Bun or Node.js required to run them.

Cache location

By default reads ~/.cache/huggingface/hub. Respects environment variables:

Variable	Description
HF_HUB_CACHE	Override cache directory directly
HF_HOME	Override HuggingFace home (cache will be $HF_HOME/hub)

Keyboard shortcuts

List view

Key	Action
↑ / ↓ or j / k	Navigate
Enter	Open detail view
d	Delete selected repo
n	Open download panel
/	Search by name
1 2 3 4	Filter: All / Models / Datasets / Spaces
s	Cycle sort: Size → Name → Last accessed → Last modified
r	Refresh cache
q	Quit

Detail view

Key	Action
↑ / ↓ or j / k	Navigate revisions
f	Toggle file scroll mode
d	Delete selected revision
D	Delete entire repo
Esc / q	Back to list

Download panel

Key	Action
Type	Enter repo ID (e.g. meta-llama/Llama-3.2-1B)
Tab	Switch type: model → dataset → space
Enter	Start download
Esc	Cancel

Delete confirmation

Key	Action
y / Enter	Confirm delete
n / Esc	Cancel

How the HuggingFace cache works

HuggingFace stores downloaded files in a content-addressable cache under ~/.cache/huggingface/hub/:

models--meta-llama--Llama-3.2-1B/
├── blobs/          ← actual file content, named by hash (real disk usage)
├── snapshots/      ← one dir per revision, files are symlinks → blobs/
│   └── abc123.../
│       ├── config.json → ../../blobs/7cb18dc...
│       └── model.safetensors → ../../blobs/403450e...
└── refs/           ← branch/tag names → commit hashes
    └── main        ← contains "abc123..."

Multiple revisions of the same model share blobs — only changed files are stored twice. This TUI measures disk usage from blobs/ (true usage, no double-counting) and cleans up orphaned blobs when deleting a revision.

Tech stack

• Ink — React-based TUI framework
• Bun — runtime, bundler, and compiler
• TypeScript

License

MIT