Multiverse Computing published a method called Quantization-Aware Healing that produces a compressed, 4-bit model reported to outperform its full-precision original1. The technique, detailed on Hugging Face, targets post-quantization accuracy recovery, claiming the reduced-precision variant exceeds the baseline rather than merely matching it.
4-Bit Quantized Model Surpasses Full-Precision Original
Multiverse Computing's Quantization-Aware Healing method produces a compressed 4-bit model reported to outperform its full-precision original.