HuggingFace Downloader

The fastest, smartest way to download models from HuggingFace Hub

Parallel downloads • Smart GGUF analyzer • Python compatible • Full proxy support

Quick Start • Why This Tool • Smart Analyzer • Web UI • Mirror Sync • Proxy Support

Why This Tool?

Parallel Downloads

Maximize your bandwidth with multiple connections per file and concurrent file downloads:

• Up to 16 parallel connections per file (chunked download)
• Up to 8 files downloading simultaneously
• Automatic resume on interruption

Real-time progress with per-file status, speed, and ETA.

Interactive GGUF Picker

Don't guess which quantization to download. Use -i for an interactive picker with quality ratings and RAM estimates:

hfdownloader analyze -i TheBloke/Mistral-7B-Instruct-v0.2-GGUF

Interactive mode features:

• Keyboard navigation - Use ↑↓ to browse, space to toggle selection
• Quality ratings - Stars (★★★★☆) show relative quality
• RAM estimates - Know if it'll fit in your VRAM
• "Recommended" badge - We highlight the best balance (Q4_K_M)
• Live totals - See combined size as you select
• One-click download - Press Enter to start, or c to copy command

Without -i, output is text/JSON — perfect for scripts and piping to other tools.

Python Just Works

Downloads go to the standard HuggingFace cache. Python libraries find them automatically:

from transformers import AutoModel
model = AutoModel.from_pretrained("TheBloke/Mistral-7B-Instruct-v0.2-GGUF")  # Just works

Plus, you get human-readable paths at ~/.cache/huggingface/models/ for easy browsing.

Works Behind Corporate Firewalls

Full proxy support including SOCKS5, authentication, and CIDR bypass rules:

hfdownloader download meta-llama/Llama-2-7b --proxy socks5://localhost:1080

Quick Start

Try it first — no installation required:

# Analyze a model with interactive GGUF picker
bash <(curl -sSL https://g.bodaay.io/hfd) analyze -i TheBloke/Mistral-7B-Instruct-v0.2-GGUF

# Download a model
bash <(curl -sSL https://g.bodaay.io/hfd) download TheBloke/Mistral-7B-Instruct-v0.2-GGUF

# Start web UI
bash <(curl -sSL https://g.bodaay.io/hfd) serve

# Start web UI with authentication
bash <(curl -sSL https://g.bodaay.io/hfd) serve --auth-user admin --auth-pass secret

Like it? Install permanently (no sudo):

bash <(curl -sSL https://g.bodaay.io/hfd) install

By default this installs to ~/.local/bin (or ~/bin if that's already on your PATH) so no sudo prompt is needed. Pass an explicit path to override:

# System-wide install (may prompt for sudo)
bash <(curl -sSL https://g.bodaay.io/hfd) install /usr/local/bin

Now use directly:

hfdownloader analyze -i TheBloke/Mistral-7B-Instruct-v0.2-GGUF
hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF:q4_k_m
hfdownloader serve
hfdownloader serve --auth-user admin --auth-pass secret   # with authentication

Files go to ~/.cache/huggingface/ — Python libraries find them automatically.

Smart Analyzer

Not sure what's in a repository? Analyze it first:

hfdownloader analyze 

For GGUF models, you get an interactive picker (see screenshot above). For other types, the analyzer auto-detects and shows relevant information:

Multi-Branch Support

Some repos have multiple branches (fp16, onnx, flax). The analyzer lets you pick:

hfdownloader analyze -i CompVis/stable-diffusion-v1-4

Diffusers Component Picker

For Stable Diffusion models, pick exactly which components you need:

Select unet, vae, text_encoder — skip what you don't need. The command is generated automatically.

Download Features

Inline Filter Syntax

Download specific files without extra flags:

# Download only Q4_K_M quantization
hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF:q4_k_m

# Download multiple quantizations
hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF:q4_k_m,q5_k_m

# Or use flags
hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF -F q4_k_m -E ".md,fp16"

Resume & Verify

# Interrupted? Just run again - automatically resumes
hfdownloader download owner/repo

# Large (LFS) files are always SHA256-verified; --verify sets the check
# for the remaining files (none|size|etag|sha256, default size)
hfdownloader download owner/repo --verify sha256

# Preview what would download
hfdownloader download owner/repo --dry-run

High-Speed Mode

# Maximum parallelism
hfdownloader download owner/repo -c 16 --max-active 8

Storage Modes

Two modes are fully supported. Pick whichever fits your workflow — neither is going away.

Mode 1 — HuggingFace cache (default, dual-layer)

hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF

Files go into the standard HuggingFace cache so Python libraries (transformers, diffusers, huggingface_hub, llama.cpp's Python bindings, …) find them automatically — nothing to configure.

~/.cache/huggingface/
├── hub/                              # Layer 1: HF cache (Python compatible)
│   └── models--TheBloke--Mistral.../
│       ├── blobs/                    # real files, content-addressed
│       ├── snapshots/a1b2c3d4.../
│       │   └── model.gguf            → symlink to blobs/
│       └── refs/main
│
└── models/                           # Layer 2: human-readable view
    └── TheBloke/
        └── Mistral-7B-GGUF/
            ├── model.gguf            → symlink to hub/.../snapshots/...
            └── hfd.yaml              # download manifest

Layer 1 (hub/): Standard HF cache structure. Python tools just work. Layer 2 (models/): Human-readable paths via symlinks — browse your downloads like normal folders.

Windows: Symlinks need Administrator or Developer Mode on Windows, so without them hfdownloader uses hardlinks instead — real files that share the same disk space, which Python, the HF CLI and you can all use. On drives without hardlinks either (FAT/exFAT), it copies, and asks once first. Choose explicitly with --link-mode auto|symlink|hardlink|copy.

Caches written by older Windows builds (files only in blobs/) are repaired by hfdownloader rebuild.

Export: plain files from what you already downloaded

Need real files for LM Studio, Ollama or llama.cpp but already downloaded into the cache? Export them — nothing is re-downloaded; the files are copied into a folder of your choice (add --mode hardlink to share the cache's disk space instead):

hfdownloader export TheBloke/Mistral-7B-Instruct-v0.2-GGUF ~/lmstudio/models/TheBloke/Mistral-7B
hfdownloader export unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF ./qwen -F q4_k_m   # just one quant

In the Web UI, start the server with --export-dir and use Export as real files on any cached repo.

Mode 2 — Flat files in a directory you choose

If you want real files at a path of your choice — no cache, no blob hashes, no symlinks — use --local-dir (matching huggingface-cli download --local-dir):

hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF \
    --local-dir ./my-model
# files land in ./my-model/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/

Unlike huggingface-cli, the files go into an // subfolder of the directory you pass. To put a repo's files directly into a folder of your choice, download normally and use export.

This is the right mode for:

• Feeding files directly to llama.cpp, ollama, or any tool that expects a plain directory of weights.
• Windows users who don't want to enable Developer Mode.
• Sharing a model over NFS, SMB, or a USB drive — hardlinks and symlinks don't travel well; real files do.
• Air-gapped transfers and manual backups.

The v2.x-compatible spelling --legacy -o produces the exact same result and is kept permanently for script compatibility:

hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF \
    --legacy -o ./my-model

Both spellings are interchangeable; pick whichever reads better in your scripts. They are mutually exclusive on a single command line.

Manifest Tracking

Every download creates hfd.yaml so you know exactly what you have:

version: "1.0"
type: model
repo: TheBloke/Mistral-7B-Instruct-v0.2-GGUF
branch: main
commit: a1b2c3d4...
repo_path: hub/models--TheBloke--Mistral-7B-Instruct-v0.2-GGUF
started_at: 2024-01-15T10:25:12Z
completed_at: 2024-01-15T10:30:00Z
command: hfdownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF -F q4_k_m
total_size: 4368438272
total_files: 1
files:
  - name: mistral-7b-instruct-v0.2.Q4_K_M.gguf
    blob: blobs/
    size: 4368438272
    lfs: true

# List everything you've downloaded
hfdownloader list

# Get details about a specific download
hfdownloader info Mistral

Web UI

A modern web interface with real-time progress:

hfdownloader serve
# Open http://localhost:8080

Cache Browser

Browse everything you've downloaded with stats, search, and filters:

All Pages

Server Options

hfdownloader serve \
  --port 3000 \
  --auth-user admin \
  --auth-pass secret \
  -t hf_xxxxx

Mirror Sync

Sync your model cache between machines — home, office, NAS, USB drive.

# Add mirror targets
hfdownloader mirror target add office /mnt/nas/hf-models
hfdownloader mirror target add usb /media/usb/hf-cache

# Compare local vs target
hfdownloader mirror diff office

# Push local cache to target
hfdownloader mirror push office

# Pull from target to local
hfdownloader mirror pull office

# Sync specific repos only (case-insensitive substring of owner/name)
hfdownloader mirror push office --repo Llama

# Verify integrity after sync
hfdownloader mirror push office --verify

Perfect for:

• Air-gapped environments: Download at home, sync to office
• Team sharing: Central NAS with all models
• Backup: Keep a copy on external drive

Proxy Support

Full proxy support for corporate environments:

# HTTP proxy
hfdownloader download owner/repo --proxy http://proxy:8080

# SOCKS5 (e.g., SSH tunnel)
hfdownloader download owner/repo --proxy socks5://localhost:1080

# With authentication
hfdownloader download owner/repo \
  --proxy http://proxy:8080 \
  --proxy-user myuser \
  --proxy-pass mypassword

# Test proxy connectivity before downloading
hfdownloader proxy test --proxy http://proxy:8080

Supported Types

Configuration File

Save proxy settings in ~/.config/hfdownloader.yaml:

proxy:
  url: http://proxy.corp.com:8080
  username: myuser
  password: mypassword
  no_proxy: localhost,.internal.com,10.0.0.0/8

Installation

One-Liner (Recommended)

bash <(curl -sSL https://g.bodaay.io/hfd) install

That's it. Works on Linux, macOS, and WSL. Installs to ~/.local/bin by default — no sudo required. Pass an explicit path to install somewhere else:

bash <(curl -sSL https://g.bodaay.io/hfd) install /usr/local/bin   # system-wide
bash <(curl -sSL https://g.bodaay.io/hfd) install ~/bin            # custom

Or run without installing:

bash <(curl -sSL https://g.bodaay.io/hfd) download TheBloke/Mistral-7B-Instruct-v0.2-GGUF
bash <(curl -sSL https://g.bodaay.io/hfd) serve   # Web UI

Download Binary

Get from Releases:

Build from Source

git clone https://github.com/bodaay/HuggingFaceModelDownloader
cd HuggingFaceModelDownloader
go build -o hfdownloader ./cmd/hfdownloader

Docker

# Pull from GitHub Container Registry
docker pull ghcr.io/bodaay/huggingfacemodeldownloader:latest

# Or build locally
docker build -t hfdownloader .

# Run (mounts your local HF cache)
docker run --rm -v ~/.cache/huggingface:/home/hfdownloader/.cache/huggingface \
  ghcr.io/bodaay/huggingfacemodeldownloader download TheBloke/Mistral-7B-Instruct-v0.2-GGUF

Private & Gated Models

For private repos or gated models (Llama, etc.):

# Set token via environment
export HF_TOKEN=hf_xxxxx
hfdownloader download meta-llama/Llama-2-7b

# Or via flag
hfdownloader download meta-llama/Llama-2-7b -t hf_xxxxx

For gated models, you must first accept the license on the model's HuggingFace page.

China Mirror

Use the HuggingFace mirror for faster downloads in China:

hfdownloader download owner/repo --endpoint https://hf-mirror.com

Or set in config file:

endpoint: https://hf-mirror.com

CLI Reference

Full documentation: docs/CLI.md • docs/API.md • docs/V3_FEATURES.md

What's New in v3.0

Both storage modes (HF cache and flat-file --local-dir / --legacy -o) are fully supported and permanent — neither is deprecated. See Storage Modes for when to pick which.

Environment Variables

License

Apache 2.0 — use freely in personal and commercial projects.

Full CLI Docs • REST API • V3 Features • Issues