Blame for README.md

---
title: README
emoji: 🧊
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
---
# Sipsa Labs

Sipsa Labs builds SIP, the operating system for the physical world.

UltraCompress is retired. Its public materials preserve historical research; new API access and paid pilots are no longer offered.

The research results below are historical and apply to their recorded versions and evaluation conditions.

---

## Historical UltraCompress research

**Historical research tiers:**

- A **genuinely lossless archival tier** — reconstructs the original weights bit-for-bit (~1.3–1.6× smaller) for exact storage and hand-off.
- A **near-lossless 5-bit tier** — ~1% perplexity vs bf16 (lossy), ~3× smaller, for running frontier-scale models on small GPUs.

**23 architectures verified**: 22 PPL-verified end-to-end (21 transformer — 17 dense + 4 MoE — plus 1 SSM with a comparator-note caveat) and 1 Vision Transformer cosine-verified (DINOv2-Large). Dense + MoE + SSM + ViT, 0.6B to 405B. The public v0.6.27 CLI checks basic pack structure and computes SHA-256 fingerprints unless hashing is skipped. It does not reconstruct weights, reproduce the evaluations below, or automatically compare its fingerprint with a trusted reference.

### Verified 5-bit PPL ratios (sample)

*(perplexity ratio = compressed PPL / bf16 baseline PPL; FineWeb-edu held-out tail; seq_len = 1024; n = 30–50; seed = 42)*

| Model | Params | Type | PPL ratio | Notes |
|---|---:|:---|---:|:---|
| **Mixtral-8x7B** | 47B | MoE | **1.00368×** | tightest MoE result |
| Qwen3-1.7B-Base | 1.7B | dense | **1.00401×** | small-decoder record |
| Qwen3-14B | 14B | dense | **1.00403×** | 14B-class record |
| Yi-1.5-9B | 8.8B | dense | **1.00414×** | >8B record |
| Qwen3-8B | 8B | dense | **1.00440×** | 8B-class record |
| Mistral-7B-v0.3 | 7B | dense | **1.00548×** | strong dense 7B-class result |
| Phi-3-mini-4k | 3.8B | dense | **1.00262×** | tightest dense ratio (seq_len=128 caveat) |
| **Hermes-3-Llama-3.1-405B** | 405B | dense | **1.0066×** | 405B-class near-lossless on a single 32 GB GPU (streamed from disk) |
| Mamba-2.8B | 2.8B | SSM | **1.0059×** | state-space model — SSM-compatible comparator (the canonical transformer pipeline is architecture-incompatible with SSMs) |
| Llama-3.1-8B | 8B | dense | **1.0125×** | honest worst case (pipeline-specific floor, not architectural) |

Full matrix at [github.com/sipsalabs/ultracompress](https://github.com/sipsalabs/ultracompress).

### Inspect a historical pack

```bash
pip install ultracompress
hf download SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5 --local-dir ./pack
uc verify ./pack
```

Expected result category: basic pack-structure checks and a computed fingerprint, not a reconstruction proof or an automatic comparison with a trusted reference.

Without an API key, `uc try` prints a recorded reference response, not a live inference result.

---

## Historical API and services

New API access and paid pilots are no longer offered. The public repository and model cards preserve historical materials.

---

## License + IP

- **PyPI v0.6+** under BUSL-1.1 with an **Additional Use Grant**: free for sub-$1M-ARR companies, research, and individuals. Auto-converts to Apache 2.0 four years after each release.
- **v0.5.x** stays Apache-2.0 on the `legacy/0.5.x` branch.
- Codec internals are patent-protected (continuations pending).

---

## Contact

- **Project inquiries** → founder@sipsalabs.com
- **Patents / licensing** → legal@sipsalabs.com
- **Press / media** → press@sipsalabs.com
- **Security disclosure** → security@sipsalabs.com
- **General** → hello@sipsalabs.com

[sipsalabs.com](https://sipsalabs.com) · [GitHub](https://github.com/sipsalabs) · [PyPI](https://pypi.org/project/ultracompress/) · [Historical repository](https://github.com/sipsalabs/ultracompress)