Back to directory

UltraCompress

A near-lossless 5-bit transformer compression method with verifiable, reproducible reconstruction: a CLI generates text against a hosted compressed model, browses a model catalog, and verifies a downloaded pack's integrity against a published SHA-256 manifest.

UltraCompress

UltraCompress compresses a transformer model from fp16 or bf16 down to a 5-bit pack while keeping quality close to the original - the authors report a 1.0066x perplexity ratio against a streaming bf16 teacher for a 405-billion-parameter model, measured on a held-out FineWeb-Edu tail. The central claim is reproducibility: an auditor can re-derive the validated quantized artifact from the pack and verify it byte-for-byte against a published SHA-256 manifest, aimed at regulated deployments in healthcare, defense, legal and quantitative-finance settings where matching a validated artifact matters.

The public package is intentionally minimal and contains no compression or reconstruction code, since that method is patent-pending; it provides a CLI to generate text against a Sipsa-hosted compressed model, browse the model catalog with quality-ratio tiers, and verify a downloaded pack's structure and download integrity. Full reproducible reconstruction of a pack is performed by Sipsa Labs under a paid engagement rather than by the open package itself.

It installs through pip and can run a first generation with no GPU and no signup.

More in Models

See all 405
  • mini-AGI

    github.com

    A continual learning language model that trains from scratch on a single 8 GB VRAM GPU, assembling its own architecture and adapting its size to the hardware and data available rather than shipping as a fixed pretrained model.

  • An open-source contrastive-embedding pipeline that maps technical CVE vulnerability descriptions to formal cybersecurity compliance controls in NIST SP 800-53 and CMMC, published with its model generation, fine-tuning and evaluation code.

  • peekaboolean

    github.com

    A small vision-language model, fine-tuned with LoRA on a 500-million-parameter backbone, that answers typed multiple-choice, ordinal-score and yes/no questions about an image with calibrated probabilities, in about 400 milliseconds on a laptop.