Unsloth Dynamic 3.0 GGUFs are post-training quantizations (no QAT/QAD) using a refined imatrix calibration dataset and per-layer quantization selection, compatible with llama.cpp and Unsloth Desktop.

Benchmarks for Qwen3.8-27B report over 10% better top-1% accuracy at equal file size versus other providers, evaluated via Divergence-300 @32 and KL Divergence on held-out prompts. File sizes range from UD-IQ1_S at 6.2GB to UD-Q2_K_XL at 9.83GB, with quants under 8.37GB omitting the MTP module to save roughly 500MB.