there is an issue with all models, including open weights. and the open weight problem will solve itself in a decade or so. after enough experimentation, someone will put the weights on dedicated hardware, and you'll get 10x token output, than it will be viable