I am using llama.cpp version b9549 with this arguments as recommended: llama-server --temp 1.0 --top-p 0.95 --top-k 64 -hf ... Here is what I got on chessboard svg test https://www.reddit.com/r/LocalLLaMA/comments/1t53dhp/quality_comparison_between_qwen_36_27b/ google/gemma-4-26B-A4B-it-qat-q4_0-gguf:IT google/gemma-4-26B-A4B-it-qat-q4_0-gguf:IT unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL For comparison here is the old gemma4 with...