TheDrainFlorist's picture
paper v5 (10.5281/zenodo.23147587)
0f0f4e5 verified
|
Raw History Blame Contribute Delete
608 Bytes
---
title: Below Five Bits
emoji: 🧮
colorFrom: gray
colorTo: blue
sdk: static
pinned: false
license: cc-by-4.0
---
# Data-Free Vector Quantization Beats Affine Quantization at Matched Bytes Below 5 Bits
Version 5, October 2026 (doi:10.5281/zenodo.23147587; all versions:
10.5281/zenodo.22119017). Vector quantization of Qwen3.5-397B-A17B,
Qwen3.6-35B-A3B and the dense Qwen3.8-27B, measured against data-free
round-to-nearest affine quantization with every other tensor held
byte-identical. No calibration data.
The VQ builds measured in the paper are published: https://huggingface.co/TheDrainFlorist