UltraCompress is retired. Its public materials preserve historical research; new API access and paid pilots are no longer offered. Historical results below apply to their recorded versions and evaluation conditions.
Near-lossless 5-bit transformer compression (~1% perplexity vs the bf16 reference; lossy), with reproducible, cryptographically verifiable reconstruction.
Public CLI v0.6.27:
uc verifychecks basic pack structure and computes fingerprints; it does not reconstruct weights or automatically compare against a trusted reference. Without an API key,uc tryprints a recorded reference response, not a live inference result.uc catalogpreserves the historical catalog.
Hermes-3-Llama-3.1-405B compressed at 5 bpw, near-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 with reproducible, cryptographically verifiable reconstruction — a deterministic decode to the SHA-256-pinned validated artifact (a near-lossless, ~1% PPL reconstruction of the bf16 source, not a bit-identical copy of it). That is the honest contract we care about: an auditor can re-derive the validated quantized artifact from the pack and verify it byte-for-byte against the published SHA-256 manifest, as described by the historical research records; the public CLI does not perform that reconstruction. 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.
This repository preserves the research and CLI documentation. It is not an offer of a current compression service.
pip install ultracompress
uc try sipsa-qwen3-0.6bWithout an API key, this prints a recorded reference response. It is not a live inference request. Historical versions may print old signup or POC links; those offers are no longer current.
uc catalogLists the 22 PPL-verified architectures (17 dense + 4 MoE + 1 SSM with comparator-note caveat) plus 1 ViT cosine-verified (DINOv2-Large) — across 4 architecture classes — with their published PPL ratios and historical tier labels (free / request / POC). Those labels do not describe current offers.
uc try [model] print a recorded reference response when no key is set
uc catalog list the historical compressed-model catalog + tiers
uc verify <pack_dir> basic pack structure + computed SHA-256 fingerprints
uc info what this package is + links/contact
uc version print version
Without an API key, uc try prints a recorded reference response. A keyed network-call path remains in the historical source; its presence is not a current API offer or an availability guarantee.
uc verify checks a readable manifest, layer-file presence, the declared layer count when supplied as an integer, and nonempty layer files. Unless --skip-hash is used, it computes SHA-256 digests and a combined pack fingerprint. It does not reconstruct weights, measure model quality or compare with a trusted expected fingerprint. Compare a computed fingerprint with a trusted published reference separately.
The default output displays a selection of hashes; --full displays all computed hashes. Both modes hash all selected pack files. --skip-hash skips fingerprint generation.
hf download SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5 --local-dir ./pack
uc verify ./packExpected result category: basic pack-structure checks and, unless skipped, a computed fingerprint. This is not a reconstruction proof or a comparison with a trusted expected hash.
The historical CLI's success text may say "download integrity verified"; computing a fingerprint alone does not establish that comparison. uc audit produces a structural/hash receipt, not a reconstruction test. See the receipt schema.
22 architectures independently PPL-verified end-to-end (0.6B → 405B, 17 dense + 4 MoE + 1 SSM) against each model's own bf16 baseline on the FineWeb-edu held-out tail at seq_len=1024, seed=42 — plus 1 Vision Transformer cosine-verified (DINOv2-Large, 304M, ViT-L/14), bringing the catalog to 23 architectures across 4 classes. 21 are transformer (17 dense + 4 MoE); the 22nd PPL-verified row is Mamba-2.8B (state-space model) at 1.00593× canonical PPL, with an explicit comparator-note caveat in the registry: our canonical transformer pipeline (RoPE / attention masks / KV-cache semantics) is architecture-incompatible with SSMs, so the Mamba record uses an SSM-compatible comparator that matches what's in the HF pack. The 23rd (DINOv2-Large) uses CLS-token cosine similarity instead of PPL (encoder-only ViT has no autoregressive likelihood). TinyLlama-1.1B (1.003×) and Llama-3.1-70B (1.009×) graduated to PPL-verified in v0.6.23. DeepSeek-32B and other queued packs remain SHA-256-verified pending canonical re-eval before formal registry entry. Every published number traces to a published result JSON. A small set of packs is publicly downloadable; the table records historical research availability, not a current service offer.
| Model | Params | Class | PPL ratio | HF artifact | Status |
|---|---|---|---|---|---|
| Hermes-3-Llama-3.1-405B | 405B | 405B-class near-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 near-lossless 5-bit ratio we currently publish.
- SSM result: Mamba-2.8B compressed with reproducible, SHA-256-verifiable reconstruction — first public near-lossless 5-bit canonical-PPL result on a state-space model that we know of, at 1.00593× canonical ratio. Counted as the 22nd PPL-verified architecture with an explicit comparator-note caveat in the registry: our canonical transformer pipeline (RoPE / attention masks / KV-cache semantics that don't apply to SSMs) is architecture-incompatible, so the Mamba record uses an SSM-compatible comparator that matches what's in the HF pack. The comparator is documented in the registry.
- HuggingFace: a small public verification set under
huggingface.co/SipsaLabs; historical catalog and artifacts. - PyPI: pypi.org/project/ultracompress.
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 pack. Mamba-2.8B ships + SHA-256-verified + canonical PPL claimed at 1.00593× (with comparator-note caveat documented in registry; our canonical transformer pipeline is architecture-incompatible with SSMs). We tried two tighter paths on top — both made it worse.
- Qwen3-32B PPL ratio. Stale or suspect baseline PPL number we won't republish. Apples-to-apples re-eval is 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.
Every other 4–5 bit compression library targets a quality threshold ("sub-1% PPL on WikiText"). UltraCompress targets a reproducible, cryptographically verifiable reconstruction contract: the published pack decodes deterministically to the SHA-256-pinned validated artifact (a near-lossless, ~1% PPL reconstruction of the bf16 reference), verifiable byte-for-byte against its manifest. 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 reproducible, verifiable deployment 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 an alternative discussed in the historical research.
As of mid-2026 we are not aware of another published library targeting a reproducible, cryptographically verifiable 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.
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.
These are historical research findings; no new design-partner engagement is offered here.
The historical 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 founder@sipsalabs.com.
- GitHub Sponsors — github.com/sponsors/sipsalabs.
- Press / commentary —
press@sipsalabs.com.
- 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.
@software{sipsa_ultracompress_2026,
author = {{Sipsa Labs, Inc.}},
title = {UltraCompress: Near-Lossless 5-bit Transformer Compression},
year = {2026},
url = {https://github.com/sipsalabs/ultracompress}
}- General:
founder@sipsalabs.com - Security:
security@sipsalabs.com - Press:
press@sipsalabs.com - HuggingFace:
huggingface.co/SipsaLabs - PyPI:
pypi.org/project/ultracompress