New method makes 4-bit AI models outperform their larger originals

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

Source: huggingface.co ↗