MLX Model Download Is Slow

First classification takes time because MLX model is downloading.

What’s Happening

On first use, para-files downloads the embedding model (~100MB) from Hugging Face:

First classification:  30-60 seconds (downloads + loads model)
Second classification: <1 second (cached)

This is normal behavior.

Speed It Up

Check Download Progress 

# First classification (shows download)
uv run para-files classify test.pdf

# Watch ~/Library/Caches/huggingface/ directory
ls -lah ~/Library/Caches/huggingface/

# Should see nomic-embed-text-v1.5

Pre-Download the Model

If you want to download before using:

# Pre-download programmatically
from para_files.encoders import MLXEncoder

encoder = MLXEncoder()
# This loads/downloads the model

# Now first classification will be fast

Or just run classify once and let it download.

Network Issues

If download is very slow:

# Check internet connection
ping huggingface.co

# Try again later
uv run para-files classify test.pdf

After First Time

Once downloaded and cached:

  • Classification: <1 second
  • Batch of 100 files: ~2-3 minutes
  • Model size: ~100MB (one-time)

Cache location: ~/.cache/huggingface/ (~100MB permanent)

Offline Use

Model works offline after first download. No network needed for subsequent classifications.