Skip to main content

Downloading Data & Model Weights

Downloading multi-gigabyte models and datasets over home broadband can take hours. Inside NovaPod instances, datacenter network backbones provide 10 Gb/s to 40 Gb/s download speeds, enabling you to pull large weights in minutes.

1. Downloading from Hugging Face​

Every NovaPod template comes with huggingface_hub pre-installed:

# Authenticate if downloading gated models (e.g. Llama 3)
huggingface-cli login --token hf_your_token

# Download an entire model repository into local directory with multi-threading
huggingface-cli download meta-llama/Meta-Llama-3-8B-Instruct \
  --local-dir /workspace/models/llama-3-8b \
  --local-dir-use-symlinks False

For ultra-fast parallel chunked downloads:

pip install hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1

huggingface-cli download Qwen/Qwen2.5-72B-Instruct --local-dir /workspace/models/qwen-72b

2. Downloading from Civitai (Stable Diffusion / Flux LoRAs)​

# Using curl with your Civitai API key
curl -L -H "Authorization: Bearer civitai_token" \
  "https://civitai.com/api/download/models/12345" \
  -o /workspace/ComfyUI/models/checkpoints/flux1-dev.safetensors

3. Downloading from Kaggle​

export KAGGLE_USERNAME="your_username"
export KAGGLE_KEY="your_api_key"

kaggle datasets download -d username/dataset-name -p /workspace/data --unzip