Unsloth's Dynamic 3.0 GGUFs squeeze 10% more accuracy out of the same quant sizes
Unsloth shipped Dynamic 3.0 GGUF quants for Qwen3.8-27B, claiming >10% better top-1 accuracy at the same size versus other providers, plus sub-7GB 1-bit quants that keep ~72% accuracy. It's all post-training quantization (no QAT/QAD), with the imatrix released and validated against 300 held-out terminal-bench, math, and multilingual prompts instead of just one-token top-1. Works with llama.cpp and Unsloth Desktop.