[Proposal] Native bounce compression for Hugging Face Hub — 25% bandwidth savings for model downloads

Research
1

Problem Statement

Downloading large AI models (10–100 GB) from Hugging Face Hub is:

  • Time-consuming for users on slower connections
  • Expensive in terms of bandwidth costs for both HF and users
  • Storage-intensive for Hugging Face infrastructure

Current compression options (gzip, zstd) are not optimized for neural network weights (IEEE-754 float tensors).

Proposed Solution

Integrate bounce compression natively into Hugging Face Hub.

Key Benefits

  • 25% average compression on model weights (.safetensors, .pt, .gguf)
  • 1069 MB/s decompression speed — faster than most network connections
  • Specialized for ML: byte-shuffle transform optimized for IEEE-754 tensors
  • CRC-32 integrity verification built-in
  • Zero dependencies: pure Rust, Apache-2.0 license

Benchmark: Safetensors Model Weights (255.5 MB)

Tool Compressed Size Ratio Decompress Speed
bounce -2 218.1 MB 85.3% 1069.0 MB/s
zstd -3 235.3 MB 92.1% 1121.8 MB/s
gzip -9 235.6 MB 92.2% 492.9 MB/s
brotli -q 5 235.1 MB 92.0% 212.6 MB/s

bounce saves 17.2 MB (7% better) than the next best tool while maintaining 5x faster decompression than gzip.

Proposed Integration

CLI

# Download with automatic decompression
huggingface-cli download model/name --compress bounce

Python SDK

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="model/name",
    filename="model.safetensors",
    compression="bounce"  # auto-decompress .bnc files
)

ROI for Hugging Face

  • Storage Savings: 25% reduction across millions of models (1 PB → 250 TB saved)
  • Bandwidth Savings: 25% less egress traffic, significant CDN cost reduction
  • User Experience: Faster downloads worldwide, lower data costs for metered connections

Resources

Open Questions

  1. Should this be opt-in or automatic for large files?
  2. Backward compatibility strategy for existing downloads?
  3. Integration timeline with huggingface_hub Python package?

I am happy to collaborate on implementation — bounce is production-ready, well-tested, and designed specifically for this use case.