ultracompress 0.1.0

Extreme compression for large language models. Download pre-compressed models from Hugging Face Hub; self-compress support coming soon.

UltraCompress

Extreme compression for large language models. Patent pending — USPTO 64/049,511 + 64/049,517

Run large language models on less hardware. UltraCompress compresses modern transformer LLMs by 26–734× with minimal quality loss. The underlying methods are patent pending; this CLI lets you download pre-compressed reference models and run them locally.

Install

pip install ultracompress

Quickstart

# List pre-compressed models available on the official Hugging Face Hub
uc list

# Download a pre-compressed model (Qwen3-1.7B at 2.798 bpw, ~30% smaller than bnb-nf4)
uc pull sipsalabs/qwen3-1.7b-uc2p79

# Inspect what's in a compressed artifact
uc info ./models/qwen3-1.7b-uc2p79

# Benchmark the compressed model against the fp16 teacher
uc eval ./models/qwen3-1.7b-uc2p79 --tasks hellaswag --limit 500

What's available today (v0.1 — alpha)

• uc list — browse pre-compressed models from our Hugging Face Hub collection.
• uc pull <model-id> — download a pre-compressed model locally.
• uc info <path> — inspect the compression metadata of an artifact.
• uc eval <path> --tasks <list> — run downstream benchmarks via lm-eval-harness on the compressed model.

What's coming

• uc compress <hf-model-id> --bpw 2.8 — self-compression (private method, coming once patents clear).
• uc serve <path> — inference server with OpenAI-compatible API.
• uc export --format gguf — export to llama.cpp GGUF format.
• uc export --format coreml — export to Apple CoreML for on-device inference.

Why UltraCompress

On a 6-model × 8-method × 500-sample head-to-head benchmark:

Method	Bits per weight	Cohort median T1 retention	Catastrophic failures
bitsandbytes int8	8.000	99.75%	0/6
bitsandbytes nf4	4.000	98.31%	0/6
HQQ 4-bit g64	4.500	97.72%	0/6
UltraCompress 2.8 bpw	2.798	95.63%	0/6
HQQ 3-bit g64	3.500	72.46%	1/6
HQQ 2-bit g64	2.500	3.46%	6/6

UltraCompress is the only sub-3-bpw method on this cohort that produces zero catastrophic failures.

Patent status

The UltraCompress compression methods are the subject of pending U.S. patent applications. Pre-compressed models are distributed under a separate licensing arrangement described in LICENSE. The CLI code in this repository is Apache-2.0.

Citation

@misc{ounnar2026ultracompress,
  title   = {UltraCompress: Extreme Compression for Large Language Models},
  author  = {Missipssa Ounnar},
  year    = {2026},
  howpublished = {\url{https://mounnar.vercel.app}}
}

Author

Missipssa Ounnar — mounnar.vercel.app · github.com/mounnar

Built on a dual-RTX-5090 workstation I designed and assembled myself. Patent pending.