YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Sipsa Labs, Inc. update — 2026-05-11. UltraCompress v0.6.2 on PyPI under BUSL-1.1 + Additional Use Grant (free for sub-$1M ARR companies, research, and individuals; auto-converts to Apache 2.0 four years post-release). OpenAI-compatible inference at api.sipsalabs.com. 22 architectures verified, 0.6B–405B parameters, sub-1.005× perplexity ratio on Mixtral-8x7B / Qwen3-14B / Mistral-7B. Live discussion on Hacker News. Commercial inquiries: founder@sipsalabs.com.
license: apache-2.0 library_name: transformers tags: - compression - ultracompress - streaming-compression - sipsa-labs
Gated artifact. This compressed model is part of Sipsa Labs' production tier (>10B parameter class). Click "Request access" to submit your use case — Sip approves manually within 24h. Free for sub-$1M ARR companies, research, and individuals (BUSL-1.1 Additional Use Grant). Above $1M ARR shipping in production: see pricing or email founder@sipsalabs.com.
phi-3.5-moe-instruct-streaming-bpw5
Compressed via UltraCompress end-to-end pipeline (per-layer streaming + correction overlay). Single 32GB GPU. Mean cross-architecture PPL_r in 9-arch matrix = 1.0066.
See compression_report.json for baseline / compressed / PPL_r details.
Reproduce
# trainer-side flags (--rank, --block_size, --train_steps) are NDA-gated; contact founder@sipsalabs.com for trainer access.
pip install ultracompress
uc compress --hf-id ... --bpw 5 --rank <r> --output ./out.uc
Patent provisionals: USPTO 64/049,511 + 64/049,517 (filed 2026-04-25).
Codec internals + training procedure are patent-protected (USPTO 64/049,511 + 64/049,517).