The paper proposes Quantization-Aware Healing (QAH), which recovers accuracy lost when a large language model is both structurally pruned and quantized to 4 bits, by distilling the compressed model directly from its original, uncompressed version rather than relying on standard quantization-aware training. Applied to a 120B model pruned to 60B parameters and quantized to MXFP4, the resulting model — Hypernova-60B — matches or beats its full-precision bfloat16 source on 7 of 9 benchmarks while using a fraction of the memory. The authors report QAH converges roughly seven times faster than baseline QAT and remains stable through continued training without manual early-stopping intervention.
