Metric	Value
PPL ratio (vs bf16 baseline)	1.0059x
Effective bpw	~5.1
Peak VRAM (single-GPU eval)	~2.5 GB
Architectures supported	22 (this is one of them)
Verifier hash	SHA256_MANIFEST.json

PPL is measured on a held-out FineWeb-edu corpus, n=30 prompts of 1024 tokens each, against the bf16 reference. Methodology details and reproducible measurement scripts ship with the verifier — see below.

This artifact is the first publicly released 5-bit near-lossless (~1% perplexity; lossy) compression of a state-space model (SSM) class architecture in the Sipsa Labs matrix.

Verify the reconstruction yourself

bash

pip install ultracompress
uc verify SipsaLabs/mamba-2.8b-hf-uc-v3-bpw5
# SHA-256 match: True
# PPL ratio (vs base): 1.0059x

The verifier downloads this artifact, decompresses to an HF folder, and checks that every weight tensor is np.allclose to the original at float32 precision. The script is open under the BUSL-1.1 Additional Use Grant.

When to use this

You want a Mamba-2.8B-class state-space model that fits in ~1.8 GB instead of the original ~5.6 GB
You need reproducible, cryptographically verifiable reconstruction (a deterministic decode to the SHA-256-pinned validated artifact), not opaque "approximately equivalent" PTQ
You're integrating into vLLM / TGI / sglang / your own stack and want a drop-in HF folder after uc unpack

When NOT to use this

You need raw fp16/bf16 — fall back to state-spaces/mamba-2.8b-hf
You're on hardware without 5-bit unpacking support — currently CUDA / Apple Metal / x86 with AVX2

License

This artifact: BUSL-1.1 with Additional Use Grant for research, individuals, and companies under $1M ARR
Auto-converts to Apache 2.0 four years after release date
Above $1M ARR running this in commercial production: founder@sipsalabs.com
Patent posture: U.S. provisional patent applications on file (April 2026); identifiers under NDA

Citation

bibtex

@software{sipsa_ultracompress_2026,
  title  = {UltraCompress: near-lossless 5-bit transformer compression at scale},
  author = {Sipsa Labs, Inc.},
  year   = {2026},
  url    = {https://github.com/sipsalabs/ultracompress},
  note   = {U.S. provisional patent applications on file; identifiers under NDA}
}

Contact

Commercial license: founder@sipsalabs.com
Security: security@sipsalabs.com
General: founder@sipsalabs.com

This artifact is part of the SipsaLabs 22-architecture matrix. Browse the full collection on the org page.