Unsloth released Dynamic v3.0, a post-training quantization method for large language models that claims >10% better top-1% accuracy than competitors at the same file size. The core innovation is a refined calibration dataset (from diverse sources, refined for coding/chat/multilingual) plus improved layer-by-layer quantization selection. They benchmarked on Qwen3.8-27B using KL Divergence (output distribution matching) and a custom "Divergence-300 @32" metric (32-token greedy decoding on 300 held-out examples) to measure both quality and overfitting. Key claimed results: the UD-Q2_K_XL variant (9.83GB) achieves ~8% higher top-1% accuracy than next-best alternatives, while UD-IQ1_S (6.2GB) retains 72% top-1% accuracy but is 89% smaller. The quants work with llama.cpp and Unsloth Desktop. The source doesn't specify whether Dynamic v3.0 is production-ready generally or still limited to Qwen—that distinction matters for practitioners deciding whether to adopt.