Quantization-Aware Healing Distills a 4-Bit Student From the Uncompressed Teacher and Beats bfloat16 on 7 of 9 Benchmarks

arXiv 2608.20953·high signalXBlueskyLinkedInCopy link

Quantization-Aware Healing (arXiv 2608.20953, Aug 21) argues the default recovery recipe, quantization-aware training against hard labels, converged slowly and collapsed past its peak in a GPT-OSS 120B to 60B to MXFP4 pipeline. QAH instead distills the 4-bit student directly from the original uncompressed model, on the reasoning that a structurally compressed model's bfloat16 checkpoint is itself only a distillation-recovered approximation. The student matches or beats its bfloat16 source on 7 of 9 benchmarks at roughly 4x less weight memory and half the teacher's parameter count, reaches a comparable peak about 7x faster than matched QAT, and is released open-weight as Hypernova-60B.