MagicQuant Qwen3.8 27B GGUFs with Unsloth v3 & Imatrix

I just updated MagicQuant with the new Unsloth dynamic v3 and utilizing their imatrix: HuggingFace Repo Here Qwen3.8 27B was a more conservative run with some kld and size savings. But it was super interesting seeing the run between dynamic v2 vs v3. The V3 was a pretty substantial boost for sure. If you don't know what MagicQuant is. TLDR, it is a benchmark driven GGUF evaluation and hybrid discovery system. It digests and learns from GGUF models like Unsloth and learns their tensor configurations. It then groups these patterns and builds isolated tests, then begins a prediction phase, and goes through a whole sha bang worth of things to try and find interesting potential hybrids. Sometimes hybrids are really crazy. Sometimes it's boring. Depends on the model and the search. If you want to really dive into the jist of MagicQuant you can learn more on my wiki here: github.com/magiccodingman/MagicQuant-Wiki Anything appended with "MQ" means it's a MagicQuant mix. But the Qwen3.8 27B results was really interesting with: Name Provider KLD Size (GB) LM-Q8_0 llama.cpp 0.000712 29.05 MQ-Q6_K_1 MagicQuant 0.000703 29.03 MQ-Q6_K_2 MagicQuant 0.000873 27.26 MQ-Q6_K_3 MagicQuant 0.001047 25.94 UD-Unsloth-UD-Q6_K_XL Unsloth 0.001238 25.33 UD-Unsloth-UD-Q6_K_L Unsloth 0.001439 24.23 MQ-Q6_K_4 MagicQuant 0.001518 23.21 MQ-Q5_K_1 MagicQuant 0.002427 22.00 MQ-Q5_K_2 MagicQuant 0.003146 20.91 MQ-Q5_K_3 MagicQuant 0.003562 20.10 MQ-Q5_K_S_1 MagicQuant 0.004818 18.98 MQ-Q4_K_M_1 MagicQuant 0.007412 17.62 UD-Unsloth-UD-Q4_K_M Unsloth 0.011205 16.50 MQ-IQ4_XS_1 MagicQuant 0.013723 16.34 UD-Unsloth-UD-Q4_K_S Unsloth 0.014502 15.39 UD-Unsloth-UD-IQ4_XS Unsloth 0.020127 14.29 UD-Unsloth-UD-Q3_K_XL Unsloth 0.030770 13.18 UD-Unsloth-UD-IQ3_S Unsloth 0.046499 12.08 MQ-IQ2_M_1 MagicQuant 0.057811 11.96 UD-Unsloth-UD-IQ3_XXS Unsloth 0.070084 10.97 MQ-IQ2_M_2 MagicQuant 0.092394 10.77 UD-Unsloth-UD-Q2_K_XL Unsloth 0.105510 9.86 UD-Unsloth-UD-IQ2_XXS Unsloth 0.182364 9.05 MQ-IQ2_XXS_1 MagicQuant 0.270304 8.27 But as always, my huggingface repo aims to show everything. The manifest folder should include all of the benchmarks results, full clone configs for replication, ran conclusions, and so on. Usually the repo is completely automated when uploaded, but this was was a fun mix so I did more manual lifting on this one to also be transparent where things came from, the dynamic v2 vs v3, and going over imatrix measurements as not the same Imatrix was used per quant. Oh and as usual, any GGUF on the survivor board that MagicQuant didn't build a mix of, the repo links back to external providers (aka: Unsloth). Anyways, this was a fun one to see imo. Unsloth kicking butt as always.