A new DeepThink model built on GLM is previewed, showing that pushing reasoning capabilities in open models could narrow the gap with costly proprietary models like Astra and Fable.