OverCompressing: A Lossless Archiver That Shrinks Fine-Tune Deltas to 0.1%
overdoingism · reddit · 2026-10-23
Built with Claude's help, OverCompressing is a lossless archiver for model folders (safetensors, GGUF, etc.), with every file verified bit-for-bit via SHA-256 on restore.
How it compresses
- Float weights (BF16/F16/F32/FP8): exponent bits are split into their own byte plane and entropy-coded while mantissa bytes stay as-is—the byte-grouping idea from IBM's ZipNN paper.
- Fine-tunes: stored as per-tensor deltas against an auto-selected base model (matched by tensor name, so sharding layout doesn't matter); the base acts as predictor for a small context-modelled entropy coder.
- Identical files (tokenizers, configs, re-uploads) stored once.
Verified results: a BF16 model compresses to 66.4% (zstd -19: 75.9%); Instruct-as-delta of Base reaches 44% (0.5B) / 30% (7B); community fine-tunes hit 0.1–5%; a 34.3 GiB test archive of six Qwen2.5 repos → 17.2 GiB (50.1%). Already-quantized GGUFs barely benefit (95–99%), though quantized fine-tunes against quantized bases reach 65%.
Formats: pack folders for upkeep with cross-model file sharing, or read-only .oczip files for cold storage. Pure Python 3.13 + NumPy + zstandard, no GPU needed; packing a 7B BF16 model takes 1.5 min, restoring 30s. MIT-licensed.
More from Infra
- Google Cloud GA's Anthropic-managed Claude deployment with day-zero model access — rseroter · 2026-10-23
- Claude Models on Google Cloud (Anthropic managed) is now generally available — claudeai · 2026-10-23
- Scaling GLM-5.3-Flash training to 1M token context, 64x over Megatron baseline — AndrewLBeam · 2026-10-23
- Perplexity's Portable Computer lands on AMD Ryzen AI Max, running local 27B model — AravSrinivas · 2026-10-23
- Liquid AI Launches RIPL: Ultra-Low-Latency Inference Engine With Sub-50ms End-to-End Responses — JosephJacks_ · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23