The user reports inconsistent performance and visual glitches when using Qwen 3.8 Flash Next on a 4×RTX 3090 setup, despite widespread praise for the model. They have experimented with various quantization levels and generated a comparison page via the 27B model to evaluate outputs. Seeking community advice on optimal quant settings to improve Flash Next reliability.