4-bit model beats its own bfloat16 original after compression
Standard quantization-recovery methods fail once a model has been structurally compressed, not just quantized: distilling from the recovered checkpoint caps accuracy at that degraded target's ceiling. QAH skips the recovered model and distills straight from the original pre-compression teacher, even across architectures. Applied to a GPT-OSS 120B compressed to 60B and quantized to MXFP4, the result beats its own full-precision bfloat16 version on 7 of 9 benchmarks. Smaller, cheaper, and more accurate than what it shipped from.