Comparison was done on Strix Halo 128gb shared memory, Ubuntu 26.04, Lemonade Server, Vulkan backend. I often run larger models like gpt-oss 120B or qwen but their performance seems to degrate quickly once in deep waters... ah.. deep context. The most important quality to me is prompt processing - we are talking existing code and context quickly fills up when analyzing it for a change request / bugfix. In existing code, I think 95-99% is PP and 1-5% is TG of the total time. I tried Nemotron Su...