Reproduce “WT-Engine and .wtq1 v2: Sub-4-Bit (3.76 bpw) Weight Quantization With 8-Bit (GGUF Q8_K) Fidelity on Qwen3-8B” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jun Hyuk Lim
WT-Engine compresses Qwen3-8B linear weights using a 3.76-bit sign-and-level carrier plus a low-rank correction that preserves information beyond calibration-activation directions. Its 3,110-MB representation reports prompt KL divergence of 0.0006 nats and 1,000-token rollout KL divergence of 0.0073 nats, with 96.7% Top-1 agreement with FP16. The rollout KL exceeds Q8_K’s 0.0020 nats, so these Qwen3-8B tests do not establish general equivalence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.