README.md

metadata

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

Sipsa Labs, Inc.

Sipsa Labs, Inc. is an experimental and deep tech-and-software company (Delaware C-corp, incorporated May 2026). We invent and ship across the full breadth of tech and software — deep research, runtime systems, novel substrates, infrastructure, hardware-adjacent stacks, and software products that don't fit anywhere else yet. UltraCompress — lossless 5-bit transformer compression — is our first flagship publicly-shipped product. More products in flight.

---

2026-05-11 — 22 architectures verified at 5 bpw, three new sub-1.005× records this week

Live on Hacker News today: https://news.ycombinator.com/item?id=48099107

Production-grade lossless 5-bit transformer compression across 22 architectures, dense + Mixture-of-Experts + state-space (Mamba), 0.6B to 405B parameters. Mathematically lossless customer-side reconstruction: SHA-256 over reconstructed tensor bytes matches the trainer'''s measurement, 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
Phi-3-mini-4k-instruct	3.8B	dense	1.00262×	seq_len=128 caveat
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 this week — 9.16× tighter than prior
Phi-3-mini-4k	3.8B	dense	1.00624×	cross-arch confirm at seq=1024
Hermes-3-Llama-3.1-405B	405B	dense	1.0066×	largest dense 5-bit lossless on the Hub
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	1.0119×	first published 5-bit lossless on a state-space model
Llama-3.1-8B	8B	dense	1.0125×	standard eval

Quick start (3 commands, no GPU required for verify)

pip install ultracompress
hf download SipsaLabs/qwen3-8b-uc-v3-bpw5 --local-dir ./qwen3-8b
uc bench ./qwen3-8b

Inference API

api.sipsalabs.com API is in private beta, OpenAI-compatible:

export OPENAI_BASE_URL=https://api.sipsalabs.com/v1
# The official `openai` SDK works unchanged.

Pricing tiers + Compression-as-a-Service contracts at https://sipsalabs.com/pricing.

License + IP

• PyPI ultracompress v0.6+ under BUSL-1.1 with Additional Use Grant — free for sub-$1M ARR companies + research + individuals. Auto-converts to Apache 2.0 four years after each release.
• v0.5.x stays Apache-2.0 forever on legacy/0.5.x.
• Two USPTO provisionals filed 2026-04-25: 64/049,511 + 64/049,517.

Commercial inquiries: founder@sipsalabs.com

Public verifier dashboard: https://sipsalabs.com/inference
Selective Disclosure Charter — what we publish vs what we keep internal: see github.com/sipsalabs/ultracompress