HuggingFace download is extremely slow or stalls

HF's CDN has regional inconsistencies — some regions get throttled to 1-5 MB/s, especially during peak hours or for new mega-models that haven't been cached locally yet. Asia and parts of Africa hit this most.

Use the HF Xet protocol (faster than the default Git LFS for new repos):

hf download <model> --xet

Try a mirror. HF-Mirror.com is a Chinese-friendly proxy:

export HF_ENDPOINT=https://hf-mirror.com
hf download meta-llama/Llama-3.1-8B-Instruct

Use Ollama's curated mirror for popular models (often faster + auto-quantized):

ollama pull llama3.1:8b

Resume failed downloads. hf download resumes automatically if interrupted. If it stalls but doesn't error, kill with Ctrl-C and re-run — it picks up from the last completed shard.

Increase parallel download workers:

export HF_HUB_DOWNLOAD_TIMEOUT=300
hf download <model> --max-workers 8

Off-peak hours (your local 3-7 AM) typically saturate available bandwidth. If you're on a metered connection at home, consider running pulls overnight.