Deep Principle has introduced Hydrogen as a model for scientific work and released its weights. Built on GLM, the system received further training using scientific knowledge and research records. The release offers downloads and hosted access, while the company’s own evaluations describe provisional research capabilities rather than independently verified discoveries.
Deep Principle, a company developing artificial intelligence for scientific research, has introduced Hydrogen-1.0 and released its weights. The model builds on Z.ai’s GLM-5.3 foundation. It is intended for scientific tasks rather than a general conversation service alone. The release combines an existing model foundation with further training using scientific knowledge, feedback from researchers and records of scientific work.[1], [2]
Deep Principle says its training material draws on chemistry, materials science and life sciences. Researchers’ feedback and records include observations, outcomes, unsuccessful attempts and revised ideas. The company presents this broader research history as a way to connect scientific knowledge with the context in which researchers work. It has published its own evaluations of scientific reasoning and other research tasks; those capability claims remain company assessments awaiting independent validation.[1]
The weights are distributed under the MIT license through Hugging Face and ModelScope, platforms that host machine-learning models. The hosted access is also described through NeuroSpark and through Mira, Deep Principle’s scientific-assistant product. The complete training dataset and training recipe do not accompany the weight release, and the company’s code and documentation have separate licensing terms.[1]
