Deep Principle releases Hydrogen-1.0, a 753B open-weight science model

The company reports that Hydrogen-1.0 outperforms GPT-6-Astra on the FrontierScience-Research benchmark, scoring 38.3% compared to GPT-6-Astra’s 33.3%. It also achieves 93.0% on GPQA-Diamond, roughly matching Kimi-K3. However, Deep Principle notes that the model trails GPT-6-Astra by 3.33 points on NMR structure elucidation tasks. The release aims to provide a reproducible and inspectable alternative to proprietary models for scientific problem-solving.

Deployment requires significant hardware, with testing conducted on 8 GPUs using SGLang. The model supports NVFP4 for NVIDIA B200/B300 GPUs and MXFP4 for AMD MI350X/MI355X. Deep Principle demonstrated the model’s capabilities in lab workflows, including writing Python for molecular dynamics simulations and planning chemical reactions. While full training datasets are not released, the open weights allow researchers to inspect and adapt the model for their specific scientific needs.