Skip to content

UltraCompress

Extreme compression for large language models.

PyPI Python License CI

UltraCompress produces near-lossless 5-bit transformer packs (~1% perplexity vs the bf16 reference; lossy) with reproducible, cryptographically verifiable reconstruction — a deterministic decode to the SHA-256-pinned validated artifact (a ~1% PPL reconstruction of the bf16 source, not a bit-identical copy of it) — 22 architectures shipped end-to-end, all 22 PPL-verified end-to-end against their bf16 baseline (plus 1 ViT cosine-verified).

v0.6.11

Pre-compressed reference models are published on the Hugging Face Hub. Run uc list for the live catalog at any time.

The compression methods are the subject of pending USPTO provisional applications filed in April 2026. This CLI is the open-source distribution layer; pre-compressed reference models roll out through the sipsalabs organization on the Hugging Face Hub through April–May 2026 — uc list shows the live catalog at any time.

Install

pip install ultracompress

uv add ultracompress

git clone https://github.com/sipsalabs/ultracompress.git
cd ultracompress
pip install -e ".[dev]"

60-second quickstart

# Browse the official catalog
uc catalog

# Download a pre-compressed model (via HuggingFace CLI)
huggingface-cli download SipsaLabs/<repo-id> --local-dir ./<repo-id>

# Verify download integrity against the SHA-256 manifest
uc verify ./<repo-id>

# Inspect manifest
uc info ./<repo-id>

What's in a pre-compressed artifact

Each artifact is a directory with:

  • model.safetensors — quantized weights in our compressed format
  • ultracompress.json — provenance manifest (bpw, base model ID, SHA-256 of weights, license, method version)
  • tokenizer/ — pre-loaded tokenizer matching the base model
  • LICENSE — the per-model license (research-free or commercial-paid; contact legal@sipsalabs.com)

Why we exist

The published methods most teams use (bitsandbytes, GPTQ, AWQ, HQQ) are all lossy — they drift relative to the original weights, with no verifiable contract on what you get back. UltraCompress ships a near-lossless 5-bit pack (~1% perplexity vs the bf16 reference; also lossy) with a reproducible, cryptographically verifiable reconstruction: a deterministic decode to the SHA-256-pinned validated artifact, verifiable against its manifest.

Method Bits per weight Reconstruction Catastrophic failures
bitsandbytes int8 8.000 lossy 0/6
bitsandbytes NF4 4.000 lossy 0/6
HQQ 4-bit g64 4.500 lossy 0/6
UltraCompress 5 bpw 5.000 near-lossless (~1% PPL), verifiable 0/6
HQQ 3-bit g64 3.500 lossy 1/6
HQQ 2-bit g64 2.500 lossy 6/6

22 architectures PPL-verified end-to-end against their bf16 baseline (FineWeb-edu held-out tail, seq_len=1024, seed=42), plus 1 ViT cosine-verified. Every published number traces to a JSON receipt.

Where to go next

  • First time here? Quickstart
  • Want to understand the methods? Compression methods overview
  • Need to integrate with your inference stack? Integration guides
  • Looking for a specific model? Run uc list for the live catalog.
  • Deploying in a commercial product? Email legal@sipsalabs.com.

Status

UltraCompress is public as of v0.6.11. The CLI is stable for list, pull, info, bench, verify, pack. Self-compression (uc compress <model>) is intentionally not yet shipped — it depends on the patent-pending compression methods being formally protected. Targeted release: late Q3 2026.

Stay in touch