Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original

Why the usual healing methods fall short here Our approach Results QAH against QAT, head to head What this changes in practice Making a large language model smaller almost always comes with a cost.