Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original
Hugging Face
Quantization-Aware Healing (QAH) showed that a GPT-OSS 120B model compressed to 60B parameters and quantized to MXFP4 can outperform its own recovered bfloat16 version when healed using KL distillation from the original pre-compression teacher instead of from the recovered checkpoint. It beats the 60B bfloat16 checkpoint on 7 of 9 benchmarks, including +7.4 on AA-LCR long-context reasoning and +5.6 on AIME 2025 math. This changes deployment pipelines by reducing reliance on quantization-aware training’s long, potentially unstable fine-tuning and enabling a safer, faster 4-bit recovery that also improves accuracy rather than only restoring it.