ARCHead: New LLM Output Head Quantization Method Substantially Reduces Storage with Minimal Loss

Şuayp Talha Kocabay · hf · 2026-08-06

While weight quantization significantly reduces the storage of Transformer blocks, the final language modeling head (LM-head) of LLMs often retains high precision, as naive quantization strongly perturbs the vocabulary logit distribution.

To address this, researchers introduce ARCHead, a compressor combining a quantized low-rank core, group-wise INT4 residuals, and a low-rank correction fitted in an activation-derived metric. This method stores no dense BF16 head and reduces persistent LM-head storage by 3.7-3.9x.

Experiments on Qwen3-8B-Base show ARCHead uses only 25.6% of BF16 head storage while achieving a 1.007 relative perplexity, vastly outperforming storage-matched naive INT4 (1.14-1.16). It serves as a lossless complement to existing block quantizers like AWQ.

Original post →

More from Infra

Infra channel →