Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original
13:39 · August 25, 2026 · Hugging Face Blog

Summary
Quantization-Aware Healing (QAH) addresses the recovery of large language models that have undergone both structural compression and low-precision quantization. After an architecture is reduced in size and its weights are cast to MXFP4, standard healing steps often leave performance below the level of the recovered full-precision checkpoint. QAH replaces the usual task-loss or post-compression distillation objective with direct knowledge transfer from the original, uncompressed full-precision model via KL divergence on output logits. Because the teacher’s distribution is architecture-agnostic, the size and shape mismatch between the 120 B teacher and the 60 B student does not prevent effective supervision.
Applied to a GPT-OSS 120 B model first compressed to 60 B parameters and then quantized, the resulting 4-bit model surpasses its own bfloat16 counterpart on seven of nine benchmarks. Gains are largest on long-context reasoning and mathematics, while the two cases where it trails lose by less than 1.5 points. The same 4-bit checkpoint also exceeds the original 120 B teacher on LiveCodeBench and remains within 1.6 points on GPQA Diamond, despite using roughly half the parameters and a quarter of the weight memory.
Compared with quantization-aware training under matched conditions, QAH reaches its peak accuracy in roughly one-seventh the number of steps and maintains that level without subsequent degradation. The stability arises because the fixed teacher distribution exerts no further pressure once the student matches it, whereas cross-entropy objectives continue to push toward hard labels and can erode earlier gains. For sequences up to 32 k tokens the method reuses a memory-efficient chunked KL implementation that processes one slice of the sequence at a time, keeping GPU memory requirements within practical limits.
The net result is a compressed model that is simultaneously smaller, cheaper to serve, and more accurate than the bfloat16 checkpoint from which it was derived, turning the quantization step into an additional round of distillation rather than a source of irreversible loss.
Why it matters
Directly addresses production challenges of quantization, memory efficiency, training stability, and benchmark-driven evaluation for compressed LLMs, with actionable recipes and quantitative results applicable by Dutch ML teams.








