Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
A GPT-OSS 120B compressed to 60B and quantized to MXFP4 beat its own bfloat16 compressed checkpoint on 7 of 9 benchmarks after Quantization-Aware Healing. For production, this makes 4-bit compressed models a potential quality upgrade rather than just a cost tradeoff, and suggests replacing long QAT-style recovery runs with quantization-aware healing when shipping structurally compressed LLMs.
Quantization-Aware Healing (QAH) enables a 4-bit compressed model to outperform its original 16-bit version on 7 out of 9 benchmarks, reversing the usual accuracy trade-off of quantization. This means shipping models that are both smaller and more accurate, reducing deployment costs while improving performance, eliminating the need for costly post-quantization recovery steps like QAT or QAD.