Run a prompt against a near-lossless 5-bit model. Right now.

No signup, no install, no API key. The weights serving this are stored at 5 bits per weight and reconstruct bit-for-bit to the validated artifact recorded in the SHA-256 manifest (the compressed pack, not the original full-precision model) — confirm the bytes match with one uc verify on your own machine.

SHA-256 verifiable 22 PPL-verified (17 dense + 4 MoE + 1 SSM) + 1 ViT cosine-verified Patents pending

5 bpw

storage format

SHA-256

reproducible reconstruction

1.0066×

405B headline PPL ratio

to try this page

Live output (streaming from api.sipsalabs.com)

Click Run to send a prompt to the live compressed model.

Prove it — in 30 seconds

The model behind this output is stored at 5 bits per weight. Every layer reconstructs to the exact validated-artifact bytes recorded in the manifest. Don’t take our word for it — run the SHA-256 check yourself:

pip install ultracompress
hf download SipsaLabs/qwen3-0.6b-uc-v3-bpw5 \
  --local-dir ./pack
uc verify ./pack   # SHA-256 manifest check

If a single byte drifts, uc verify fails loudly. That is the difference between “reproducible” as marketing and “reproducible” as a contract you can audit. Public artifacts → HuggingFace model
→ GitHub source
→ PyPI package Verified record (this model) Qwen3-0.6B PPL ratio: 1.0069× (+0.69%)
Eval: FineWeb-edu, n=30, seq_len=1024, seed=42
→ full benchmark matrix (22 PPL-verified + 1 ViT cosine-verified) IP Implementation details are proprietary and patent-pending. The verifier flow, binary format, and reconstruction contract are public.

Want your own key? Self-serve API access ships with $5 free credits, no card. Use the same model menu (22 PPL-verified + 1 ViT cosine-verified, OpenAI-compatible endpoint at api.sipsalabs.com/v1). → /get-access · /pricing

JavaScript is disabled. The interactive playground needs JS to stream output. To try inference without JS:

Read the verified benchmarks page
Or run it locally: pip install ultracompress && uc verify ./pack

Sample interaction (no JS required)

prompt → sipsa-qwen3-0.6b

Q. Write a haiku about a model file weighing five bits per weight.

A. Five bits per weight, light —
a thousand layers folded small,
the same words emerge.

verification

Output above produced by the same model file you can download from SipsaLabs/qwen3-0.6b-uc-v3-bpw5. SHA-256 manifest on the HF page; uc verify confirms pack structure + download integrity.