According to CoinMeta, Zhipu GLM-5.3 was taken to the laboratory for special training by scientists, and its research performance surpassed that of GPT-6 and Astra. A scientific research model hydrogen-1.0 was released, which was trained based on Zhipu GLM-5.3. In the OpenAI's Frontierscience-research scientific research test, hydrogen-1.0 scored 38.3 points, exceeding the 33.3 points of GPT-6 and Astra; the original version of GLM-5.3 only scored 24.4 points. The training materials for this model are quite unique, with complete experimental records and trial processes being used for its development. During actual demonstrations, hydrogen-1.0 was able to write code on its own, call scientific research software to simulate the movement of plastic molecules, and even correct setting errors by itself. The team also had it perform five rounds of new material screening in a row; the candidate materials found ultimately surpassed GPT-6, Astra, and Kimi-K3 according to the team's set criteria. The model weights have been made public according to the MIT protocol, but the laboratory data used for training has not been disclosed.