ultracompress 0.6.15

Lossless 5-bit transformer compression — 14 architectures independently PPL-verified end-to-end (0.6B–405B; dense, MoE, SSM). Public CLI does pack structure + download-integrity checks; the codec is patent-pending and not distributed. BUSL-1.1 + Additional Use Grant.

UltraCompress

Lossless 5-bit transformer compression. Published model artifacts are bit-identical to their bf16 reference.

v0.6.15: the public package is intentionally minimal — a small, dependency-free CLI that checks pack structure and download integrity and prints project info. It contains no compression or reconstruction code: that methodology is patent-pending and is not distributed. Bit-identical reconstruction verification of a pack is performed by Sipsa Labs under engagement. The compressed model artifacts themselves remain public on HuggingFace.

Hermes-3-Llama-3.1-405B compressed at 5 bpw lossless: 1.0066x PPL ratio vs streaming bf16 teacher (5.0692 / 5.0358, n=50, seq_len=1024, FineWeb-edu held-out tail, seed=42). A 405B-class transformer compressed end-to-end on a single 32 GB consumer GPU.

UltraCompress takes a transformer at fp16/bf16 and produces a 5-bit pack that reconstructs bit-identically to the reference bf16 checkpoint — not "1% PPL drift on WikiText," but a deterministic reconstruction. That is the honest definition of lossless we care about: an auditor can re-derive every weight from the pack, and Sipsa Labs verifies that bit-identity under engagement. The codec is patent-pending.

It exists because the bf16-equivalent quality bar matters in places where "good enough on MMLU" isn't enough — defense, FDA-regulated healthcare, SR 11-7 model validation, internal red-team eval at frontier labs. And as a side-effect of the streaming compression path, it lets us put a 405B-parameter model through a single 32 GB consumer GPU without renting an H100 cluster.

We're a small lab shipping this in public while the patents are pending. Most days the lab notebook gets longer than the marketing site does.

---

The public CLI (what pip install gives you)

pip install ultracompress
uc info

uc verify <pack_dir>   pack structure + download-integrity self-check
uc info                what this package is + links/contact
uc version             print version

uc verify confirms a downloaded pack is well-formed (manifest present and parseable, declared layer count matches the files on disk, no zero-byte layers) and prints a stable SHA-256 pack fingerprint so you can confirm you hold a byte-identical download, or compare against a fingerprint we publish out of band. It does not reconstruct weights and contains no codec knowledge by design.

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

uc_pack_version: 3
bpw:             5
layer files:     28
SHA-256 (spot-check; use --full for all):
  manifest.json:f3a1…
  layer_000.uc:7c2b…
  layer_014.uc:9d4f…
  layer_027.uc:1ab8…
pack fingerprint (sha256 of sorted file digests):
  4e9c… (64 hex)

→ STRUCTURE OK — pack is well-formed; fingerprint above is the
  download-integrity reference. This is NOT a reconstruction proof;
  bit-identical reconstruction verification is provided by Sipsa
  Labs under engagement ([e-mail]).

Full bit-identical reconstruction verification (and PPL re-evaluation against the bf16 baseline) is an auditor-grade deliverable Sipsa Labs runs with you under engagement — it is deliberately not shipped in the public package.

---

What's verified (with JSON receipts)

14 architectures independently PPL-verified end-to-end (0.6B → 405B, dense + MoE + state-space) against each model's own bf16 baseline on the FineWeb-edu held-out tail at seq_len=1024, seed=42. Every published number traces to a published result JSON. A small set of packs is publicly downloadable; the full catalog is available to customers under engagement.

Model	Params	Class	PPL ratio	HF artifact	Status
Hermes-3-Llama-3.1-405B	405B	405B-class lossless on a single 32 GB consumer GPU	1.0066	SipsaLabs/hermes-3-llama-3.1-405b-uc-v3-bpw5	live
Mistral-7B-v0.3	7.2B	sub-0.6% drift	1.00548	SipsaLabs/mistral-7b-v0.3-uc-v3-bpw5	live
Qwen3-1.7B-Base	1.7B	sub-0.5% drift	1.00401	SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5	live
Qwen3-14B	14.0B	sub-0.5% drift	1.00403	SipsaLabs/qwen3-14b-uc-v3-bpw5	live
Qwen3-8B	8.0B	sub-0.5% drift	1.00440	SipsaLabs/qwen3-8b-uc-v3-bpw5	live
Mixtral-8x7B-v0.1 (MoE)	47B (13B active)	sub-0.5% drift	1.00368	SipsaLabs/mixtral-8x7b-v0.1-uc-v3-bpw5	live
Phi-3-mini-4k-instruct	3.8B	sub-0.3% drift (seq_len=128, not apples-to-apples)	1.00262	SipsaLabs/phi-3-mini-4k-instruct-uc-v3-bpw5	live

Hermes-3-405B is the headline. The 1.0066x ratio is 5.0692 / 5.0358 — both halves measured under the same per-layer streaming reconstruction comparator (n=50, seq_len=1024, FineWeb-edu held-out tail, seed=42). The bf16 teacher took 7.7 hours on cuda:1; the 5-bpw pack took 14.3 hours. The Mistral-7B 1.00548× row is the tightest dense 7B-class lossless 5-bit ratio we currently publish.

• Lossless 5-bit state-space-model compression: Mamba-2.8B at 1.0119.
• HuggingFace: a small public verification set under huggingface.co/SipsaLabs; full catalog under engagement.
• PyPI: pypi.org/project/ultracompress.

---

What doesn't work yet

Things people sometimes assume work because the rest of it does. They don't, and we'd rather you know:

• Long-context evaluation past seq_len=1024. Every PPL number above is at seq_len=1024 on the FineWeb-edu held-out tail. We have not yet run controlled evals at 4K/8K/32K context.
• State-space models past the current SSM result. Mamba-2.8B at 1.0119 is the SSM number, full stop. We tried two tighter paths on top — both made it worse.
• TinyLlama-1.1B-Chat PPL eval. The pack itself is well-formed and the HF artifact uploaded, but the PPL eval forward pass throws a CUDA device-side assert that we haven't traced yet. Shown as deferred, not a fabricated number.
• Qwen3-32B and Llama-3.1-70B PPL ratios. Both have stale or suspect baseline PPL numbers we won't republish. Apples-to-apples re-evals are queued.
• Below 1.0040× on Qwen3-1.7B-Base. This is our tightest dense floor; we tried 5 different paths to break it. Three were within noise; two were catastrophic regressions. 1.0040× stands as the empirical floor at the current configuration.

---

Why this isn't AWQ / GPTQ / EXL3

Every other 4–5 bit compression library targets a quality threshold ("sub-1% PPL on WikiText"). UltraCompress targets a reconstruction contract: the published artifact reconstructs bit-identically to the reference bf16 checkpoint. Codec internals are patent-pending and deliberately not described here.

This matters when "the model picks a slightly-wrong variable name" is a regulatory finding rather than a cosmetic complaint. Defense / aerospace deploy-bit-exactness is a compliance requirement. FDA-regulated healthcare AI requires model equivalence between dev and deploy. SR 11-7 (Federal Reserve model validation) requires reproducible audit recovery.

For pure-throughput inference on a fixed prompt distribution that matches your AWQ calibration set, with no downstream fine-tuning, AWQ at 4 bpw on vLLM is genuinely fine and we'll say so on a sales call.

As of mid-2026 we are not aware of another published library targeting a bit-identical reconstruction contract (as opposed to a PPL-threshold) for 5-bit transformer compression on the public HuggingFace Hub. If you find one, tell us — we'd rather benchmark against it than claim a gap that isn't there.

---

Honest negative results

Most projects hide their failures. We catalogue them at the same level of detail as the wins.

• An initialization shortcut we tried — made PPL 0.07 pp WORSE on Mamba and was discarded. Method specifics withheld (patent-pending).
• A multi-pass variant we hypothesized would help — produced a catastrophic 13.7× regression vs. the single-pass baseline. CLOSED.
• Importing an AWQ-style pre-scaling step — produced a catastrophic +13% regression and was ruled out. CLOSED.
• Pushing the training schedule past the current configuration — gained nothing (within noise). The floor stands.
• "Base models compress tighter than instruct" hypothesis — refuted 2/3 of architectures. Dropped.

Detailed methodology for any specific failure is available to design partners under NDA.

---

Who this is for

• If you serve LLMs in production and your VRAM bill is the constraint, this might help. It scales to a 405B-class model on a single 32 GB consumer GPU (the how is patent-pending). Email [e-mail] with your stack and a target latency/quality bar; we'll tell you honestly whether UC fits.
• If you're in a regulated domain (defense, FDA-regulated healthcare, SR 11-7 model validation, frontier lab red-team), the bit-identical reconstruction contract is the reason to talk to us. Phase 0 POC ($5K, 5 business days, customer-picked model) gets you a pack plus a Sipsa-run bit-identity + PPL audit you can review. Email [e-mail].

If your workload is "MMLU has to stay above X" and you're not pushing the model into long-tail or downstream-fine-tuning territory, AWQ at 4 bpw is probably a better answer than this. We'll say so.

---

We're a small company looking for design partners

Sipsa Labs is a small lab shipping in public. Our compression methods are patent-pending; details are in PATENT_NOTICE.md. The CLI source is BUSL-1.1 with an Additional Use Grant — free for companies under $1M ARR, research, and individuals, auto-converting to Apache 2.0 four years post-release. If you're building a derivative product whose core value depends on the underlying invention, email [e-mail].

• Paid Phase 0 POC — [e-mail], $5K / 5 business days / customer-picked model. Deliverable: a pack plus a Sipsa-run bit-identity + PPL audit on your eval set.
• GitHub Sponsors — github.com/sponsors/sipsalabs.
• Press / commentary — [e-mail].

---

License

• Released under BUSL-1.1 with an Additional Use Grant (free for companies under $1M ARR, research, and individuals; auto-converts to Apache 2.0 four years post-release). See LICENSE.
• The license grant does not extend to the patent-pending compression methodology that produces the artifacts. See PATENT_NOTICE.md.
• Pre-compressed model artifacts on HuggingFace carry the upstream teacher model's license plus this project's patent terms.

Citation

@software{sipsa_ultracompress_2026,
  author = {{Sipsa Labs, Inc.}},
  title  = {UltraCompress: Lossless 5-bit Transformer Compression},
  year   = {2026},
  url    = {https://github.com/sipsalabs/ultracompress}
}

Contact

• Commercial / Phase 0 POC: [e-mail]
• Security: [e-mail]
• Press: [e-mail]
• HuggingFace: huggingface.co/SipsaLabs
• PyPI: pypi.org/project/ultracompress