You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

phi-3.5-moe-instruct-uc-v3-bpw5

pip install ultracompress
uc try sipsa-qwen3-0.6b      # 30s demo (free, no signup, no GPU)

This calls Sipsa's free inference API on a compressed model — see what 1.004x PPL ratio looks like in practice.

To run this model via the API: uc try sipsa-phi-3.5-moe

Browse the full catalog of try-able models: uc catalog

Phase 0 POC: $5K / 5 business days / SHA-256 reproducible-reconstruction audit (deterministic decode to the validated artifact).

Near-lossless 5-bit compression (~1% perplexity; lossy) of microsoft/Phi-3.5-MoE-instruct — produced by Sipsa Labs (UltraCompress).

Reproducible, cryptographically verifiable reconstruction: a deterministic decode to the SHA-256-pinned validated artifact (not bit-identical to the original bf16 model).
Independently perplexity-verified end-to-end.
PPL ratio vs bf16: 1.00129× (FineWeb-edu held-out tail, seq_len=1024, seed=42, n=30 prompts).
License: BUSL-1.1 + Additional Use Grant.

pip install ultracompress

Patents pending.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Base model

Finetuned

(6)

this model