File size: 5,304 Bytes
fd67247 4c127d2 fd67247 4c127d2 fd67247 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | ---
title: Sipsa Labs
emoji: 🧊
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
---
# Sipsa Labs
> **Sipsa Labs** is an experimental and deep tech-and-software company. We invent and ship across the full breadth of tech and software.
>
> **Two products live today:**
>
> 1. **UltraCompress** — lossless 5-bit transformer compression. Flagship. `pip install ultracompress`.
> 2. **Sipsa Inference** — OpenAI-compatible serving API for our compressed weights. `https://api.sipsalabs.com/v1`. Pro $99/mo + Team $499/mo.
>
> **Compression-as-a-Service** engagements are open ($5K Phase 0 POC). More products in flight — we don't pre-announce.
---
## v0.6.15 current — minimal clean release
Only `pip install ultracompress` (v0.6.15) is on PyPI. The package is the minimal SHA-256 verifier; codec internals are patent-protected and not shipped in the public package. Prior dev releases were retired during the 2026-05-18 cleanup. Mistral-7B-v0.3 hits 1.00548× — a strong dense 7B-class lossless 5-bit result.
---
## Product 1: UltraCompress
Production-grade lossless 5-bit transformer compression across 22 shipped architectures (14 PPL-verified end-to-end), dense + Mixture-of-Experts + state-space, 0.6B to 405B parameters. Mathematically lossless customer-side reconstruction: SHA-256 over reconstructed tensor bytes matches the reference manifest, verified by `uc verify`.
### Tightest verified PPL ratios at 5 bpw
*(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×** | NEW — a 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 lossless on a single 32 GB consumer GPU |
| Qwen3-0.6B | 0.6B | dense | **1.0069×** | |
| OLMo-2-0425-1B | 1B | dense | **1.0073×** | |
| SmolLM2-1.7B-Instruct | 1.7B | dense | **1.0075×** | |
| SmolLM2-1.7B | 1.7B | dense | **1.0085×** | |
| Mamba-2.8B | 2.8B | SSM | _compression validated; PPL eval pending_ | state-space model |
| Llama-3.1-8B | 8B | dense | **1.0125×** | |
22 architectures shipped, 14 PPL-verified end-to-end. Full matrix below + at github.com/sipsalabs/ultracompress.
### Try UltraCompress (3 commands)
```bash
pip install ultracompress
hf download SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5 --local-dir ./pack
uc verify ./pack
```
Expected: `VERIFY: PASS — bit-identical reconstruction guaranteed.`
---
## Product 2: Sipsa Inference
OpenAI-compatible inference API serving our compressed weights. Drop-in replacement for OpenAI's `base_url` — same `openai` Python SDK works unchanged.
```bash
export OPENAI_BASE_URL=https://api.sipsalabs.com/v1
curl $OPENAI_BASE_URL/models
```
22 models live in the catalog with `sipsa-*` prefix (e.g. `model="sipsa-hermes-3-llama-3.1-405b"`). Backed by dual RTX 5090 over Cloudflare Tunnel.
**Pricing**: Free $5 credit on signup (no card). **Pro $99/mo** (600 RPM, $100 included credit). **Team $499/mo** (2400 RPM, $500 included credit). Full pricing + bill estimator at sipsalabs.com/pricing.
---
## Service: Compression-as-a-Service (CaaS)
Bring a model, we deliver a verified-lossless 5-bit pack you can run on your hardware. **Phase 0 POC is $5K / 5 business days / customer-picked model**. Day 7 deliverable is a pack you self-verify with `uc verify` + benchmark with `uc bench`. Acceptance gate: `uc verify PASS` + PPL ratio within 1.5% on your eval set.
Email founder@sipsalabs.com.
---
## Why this matters
Defense / FDA-regulated healthcare / SR 11-7 model validation / frontier lab red-team eval — places where "good enough on MMLU" isn't enough. Bit-identical reconstruction is the actual reason to talk to us in those settings. As a side-effect of the streaming compression path, it lets us put a 405B-parameter model through a single 32 GB consumer GPU without renting an H100 cluster.
---
## License + IP
- **PyPI v0.6+** under BUSL-1.1 with **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 under Apache-2.0 forever on the `legacy/0.5.x` branch.
- Codec internals patent-protected. Continuations through 2027.
---
## Contact
- **Commercial / Phase 0 POC** → 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/) · [Pricing](https://sipsalabs.com/pricing)
|