I am running 27B with Deepseek Harness these days and somehow just by using it, without any parameter changes, the model feels even more intelligent.