Dynamic Insight Beating AI News: Deep Principle (Deep Principles) has released the research model Hydrogen-1.0, trained on Zhipu GLM-5.3. In OpenAI's FrontierScience-Research scientific evaluation, Hydrogen-1.0 scored 38.3 points, surpassing GPT-6 Astra's 33.3 points and the original GLM-5.3's 24.4 points.

Its training materials are particularly unique. Scientific papers typically only present final results, but real research involves understanding how experiments are designed, where errors occur, and what adjustments to make next. Deep Principle used its own lab's complete experimental records, along with iterative attempts and revisions, to train the model.

In practical demonstrations, Hydrogen-1.0 can autonomously write code and use research software to simulate the movement of plastic molecules. When it detects an error in simulation settings, it fixes it itself and continues running and analyzing results. In another demonstration, it devised experimental plans for 12 chemical reactions and identified in a virtual lab that the step of replacing pipette tips was missing—promptly adjusting the procedure.

The team also had it perform five consecutive rounds of new material screening. Each round adjusted based on the previous round’s evaluation results to determine the next set of trials. When searching for metal complexes with target electronic properties, it ultimately found candidate materials that outperformed both GPT-6 Astra and Kimi-K3 according to the team's criteria.

The model weights are open-sourced under the MIT license, but the laboratory data used for training remains unpublished.