Preprint improves language-model accuracy at sub-1-bit weight storage

Researchers report that ShamAN-Q, a method for compressing language-model weights after training, improves Qwen3-Base accuracy while retaining NanoQuant's deployment format. At approximately 1 bit per weight, WikiText-2 perplexity falls from 27.56 to 22.96 for the 0.6B model, from 19.21 to 16.72 for 1.7B, and from 14.29 to 13.80 for 4B. The method uses calibration data to adjust how compression errors are weighted and redistributes storage across layers without increasing total bits. On the 0.6B model, approximately 0.8 bits per weight matches NanoQuant's published perplexity at approximately 1.0 bit per weight. These are preprint results on Qwen3-Base.