Hangzhou Deep Principle has open-sourced Hydrogen-1.0, a scientific large language model with 753 billion parameters in a mixture-of-experts architecture (about 40 billion active), released under the MIT license.
Post-trained on GLM-5.3, the model targets discovery in chemistry, materials and the life sciences. The team says Hydrogen-1.0 surpasses GPT-6-Astra on the FrontierScience-Research benchmark.
The release adds to a growing wave of Chinese open-weight scientific models aimed at accelerating research workflows in the physical sciences, distributed freely for academic and commercial reuse.
Source: China News Report (compiled from Toutiao AI Industry Hotspot).
