[Zhipu GLM-5.3 Trained in Scientist Labs, Surpasses GPT-6 Astra in Research Benchmarks] Beating AI News: Deep Principle released its research model Hydrogen-1.0, trained on Zhipu GLM-5.3. In OpenAI's FrontierScience-Research benchmark, Hydrogen-1.0 scored 38.3 points, surpassing GPT-6 Astra's 33.3 points, while the original GLM-5.3 only scored 24.4 points. Its training materials are unique. Scientific papers typically only present the final results, but real research requires understanding how experiments are designed, where errors occurred, and how to adjust them. Deep Principle used its own complete experimental records, along with the process of trial and error, to train the model. In actual demonstrations, Hydrogen-1.0 can write its own code to call scientific software to simulate the movement of plastic molecules. If a simulation setting error is detected midway through, it will automatically correct itself and continue running and analyzing the results. In another demonstration, it developed experimental protocols for 12 chemical reactions and even discovered in the virtual lab that a pipetting operation had missed the step of changing pipette tips, subsequently modifying the procedure. The team also had it perform five rounds of new material screening. Each round determined what to try next based on the evaluation results of the previous round. When searching for metal complexes with target electronic properties, the candidate materials it ultimately found surpassed GPT-6 Astra and Kimi-K3 on the team's set metrics. The model weights have been released under the MIT license, but the laboratory data used for training has not been made public. [Original Link]