Skip to content
@sipsalabs

sipsalabs

Sipsa Labs

Sipsa Labs, Inc. — an experimental and deep tech-and-software company. 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.

🔥 Live on Hacker News today (2026-05-11): news.ycombinator.com/item?id=48099107


First flagship product: UltraCompress

Lossless 5-bit transformer compression. SHA-256 verifiable bit-identical reconstruction at customer load. Different contract than every other 4–5 bit library: we don't target a quality threshold ("sub-1% PPL drift"), we target a reconstruction contract — the customer artifact reproduces exactly the dequantized weight the trainer measured during distillation. If anything drifts, uc verify fails loudly.

This week's verified PPL ratios (5 bpw vs bf16, FineWeb-edu held-out tail, seq_len=1024, seed=42)

Model Class PPL ratio HF artifact
Hermes-3-Llama-3.1-405B First 405B-class lossless 5-bit on a single 32 GB consumer GPU 1.0066× SipsaLabs/hermes-3-llama-3.1-405b-uc-v3-bpw5
Mixtral-8x7B (47B MoE) Tightest MoE result 1.00368× SipsaLabs/mixtral-8x7b-v0.1-uc-v3-bpw5
Qwen3-1.7B-Base Tightest dense floor 1.00401× SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5
Qwen3-14B 14B-class 1.00403× SipsaLabs/qwen3-14b-uc-v3-bpw5
Qwen3-8B 8B-class 1.00440× SipsaLabs/qwen3-8b-uc-v3-bpw5
Mistral-7B-v0.3 New this week — 9.16× tighter than prior public best 1.00548× SipsaLabs/mistral-7b-v0.3-uc-v3-bpw5
Phi-3-mini-4k-instruct Cross-arch confirm 1.00624× SipsaLabs/phi-3-mini-4k-instruct-uc-v3-bpw5

22 architectures verified end-to-end (0.6B → 405B, dense + Mixture-of-Experts + state-space). Full matrix at huggingface.co/SipsaLabs.

Try it (3 commands)

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

Or use the OpenAI-compatible API (no install)

export OPENAI_BASE_URL=https://api.sipsalabs.com/v1
curl $OPENAI_BASE_URL/models

The official openai Python SDK works unchanged — same client.chat.completions.create(), same SSE chunks. Backed by dual RTX 5090 over Cloudflare Tunnel.

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.
  • USPTO provisionals 64/049,511 + 64/049,517 filed 2026-04-25. Supplement filing landed 2026-05-09. Continuations through 2027.

What's next

  • More products in flight across runtime systems, novel substrates, software infrastructure (announced as they ship — we don't pre-announce).
  • Continued architecture coverage on UltraCompress: 70B/235B/685B in the queue.
  • Public NeurIPS 2026 + ICLR 2027 paper drafts.

Contact

sipsalabs.com · HuggingFace · PyPI · Hacker News (live today)

Popular repositories Loading

  1. ultracompress ultracompress Public

    Lossless 5-bit transformer compression with OpenAI-compatible API. 22 architectures verified, 0.6B-405B. Mistral-7B 1.005x. Hermes-3-405B 1.0066x. pip install ultracompress

    Python 11

  2. .github .github Public

    Sipsa Labs organization profile

Repositories

Showing 2 of 2 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…