Hydrogen-1.0 Open Science AI Model Outperforms GPT-6-Astra

An open-weight AI model named Hydrogen-1.0 has been released, reportedly outperforming GPT-6-Astra in real scientific research benchmarks. Built on GLM-5.3 and trained on actual lab records and discovery steps, the model achieved 38.3% on FrontierScience-Research and 93.0% on GPQA-Diamond. Its creators note that any lab can now download the weights to run and adapt the model for their own research. Open article →