UltraCompress let you run frontier models on hardware you already own, with reproducible, cryptographically verifiable reconstruction — the model your audit reviewed is the model you ship. The mechanism is near-lossless 5-bit transformer compression (~1% perplexity): ~3× smaller weights (16-bit → ~5 bits/weight) at the same task quality. The research, benchmarks, and published packs remain available below as a record.
# Public substrate (no API key required):
pip install ultracompress
hf download SipsaLabs/qwen3-8b-uc-v3-bpw5 --local-dir ./qwen3-8b
Same 5-bit substrate, three delivery modes during the commercial run — from a free pip install on your laptop to a managed OpenAI-compatible endpoint to a deployment inside your security boundary. Today only the free, public path remains active.
"I have a 5090 / 4090 / Mac with 32 GB. Give me the weights."
Free public substrate. pip install ultracompress, hf download the artifact, run uc verify. No signup. No API key. The same SHA-256 manifest your security team would audit, on your hardware, today.
"Swap the base URL. Cut the bill. Don't rewrite the app."
The managed API operated on this stack during the commercial run; it is no longer sold. The open substrate (PyPI + published packs) remains available for researchers.
"It has to run inside our boundary, under our audit log."
Reproducible reconstruction inside the customer's VPC, bare-metal cluster, or air-gapped enclave. SOC 2 / SR-11-7 / FDA / DoD-ready architecture. This path closed when the product was discontinued.
Every record below has a public Hugging Face artifact and an SHA-256 manifest you can re-verify on your hardware. Perplexity ratio is measured against the bf16 baseline at seq_len=1024, FineWeb-edu held-out tail. Run uc verify and confirm the contract holds — no "trust me."
Four things made this different from every other "model compression" pitch. Each one remains verifiable on your hardware from the published artifacts — nothing to sign, nothing to pay.
Set OPENAI_BASE_URL=https://api.sipsalabs.com/v1 and your existing inference code keeps working. Chat, completions, embeddings — same surface area, same response shape. Drop-in replacement for the OpenAI client in Python, Node, Go, Rust.
Every artifact ships with a per-file SHA-256 manifest. uc verify confirms the downloaded bytes match that manifest (pack structure + download integrity; it does not run the model or reproduce perplexity). The compressed artifact your audit reviewed in March is the artifact your endpoint serves in October. SR-11-7 and FDA SaMD reviews carry through — no "compressed-variant" governance lane.
Task quality preserved, measured, published. Perplexity ratios from ~1.0013× to ~1.0125× against the bf16 baseline (1.0066× on the 405B flagship) — not "looks fine to me," not eyeball-tested on three prompts. Reproduce the eval on your hardware with one command. And when you need exactness over near-lossless, a separate genuinely lossless archival tier reconstructs the original weights bit-for-bit.
Fits on consumer GPUs you already own — or consolidates onto far fewer of the GPUs you already rent (one box where the bf16 weights needed several). Hermes-3-405B on a single RTX 5090. Mixtral-8x7B on a 4090. The cost-per-token math changes when the weights stop spilling into a second box.
The free path still runs end-to-end today — no signup, the substrate is public. The managed and enterprise paths closed when the product was discontinued; they stay below as a record of how it shipped.
Install the CLI, pull a published artifact, run the verifier and benchmark. Three commands. The full substrate, free to run, under BUSL-1.1 with the Additional Use Grant.
# 1. Install
pip install ultracompress
# 2. Pull an artifact (Hugging Face)
hf download SipsaLabs/qwen3-8b-uc-v3-bpw5 --local-dir ./qwen3-8b
# 3. Verify SHA-256 + run benchmark on your hardware
uc verify ./qwen3-8b
During the commercial run, the API served an OpenAI-compatible endpoint. The page below is kept as a record of what shipped.
# 1. Keys are no longer issued — purchases closed June 2026
export SIPSA_API_KEY=sk-...
# 2. Use the standard OpenAI SDK, just change the base URL
from openai import OpenAI
client = OpenAI(
base_url="https://api.sipsalabs.com/v1",
api_key=os.environ["SIPSA_API_KEY"],
)
resp = client.chat.completions.create(
model="sipsa-qwen3-8b",
messages=[{"role": "user", "content": "Hello"}],
)During the commercial run, enterprise deployments ran inside the customer's security boundary — on-prem, VPC, or air-gapped, with SOC 2 / SR-11-7 / FDA / DoD-ready architecture. This path is closed; the intake template is kept as a record.
# Archived — the enterprise intake asked for:
# Use case
# Scale (GPU count, expected token throughput, models)
# Security boundary (on-prem, VPC, air-gapped)
# TimelineUltraCompress sold as a self-serve subscription during its commercial run. The product was discontinued in June 2026 and purchases are closed — the tiers below are kept, with all purchase links removed, as a record of what shipped. The free substrate (PyPI + published packs) remains public.
For evaluation and small workloads. Real inference, real models, real SDK.
For individual developers running real workloads.
For production apps with real customer traffic. 5× or 20× the Pro quota.
For engineering teams with central billing and admin. 5 seats minimum.