Using GLM. It works but chews tokens pretty hard