Deep principle unveils Hydrogen-1.0, a 753 billion parameter open-weight model for scientific research

Deep Principle, a Hangzhou-based AI firm, has released Hydrogen-1.0, a 753B open-weight model designed to assist in chemistry, materials science, and life sciences. This model is derived from Z.ai’s GLM-5.3 and was trained using domain data, laboratory records, and agent discovery trajectories. The weights are available under the MIT License, offering an inspectable alternative to proprietary systems. Hydrogen-1.0 demonstrates strong performance against competitors. On the FrontierScience-Research benchmark, Deep Principle reports it surpasses GPT-6-Astra. Furthermore, a founder's tweet indicated the model achieved a 93.0% score on GPQA-Diamond, matching Kimi-K3. In internal evaluations, Hydrogen-1.0 achieved a mean best-so-far HOMO-LUMO gap of 4.318 eV in transition metal complex design over five rounds, outperforming GPT-6-Astra (3.689 eV) and Kimi-K3 (3.353 eV). The model showcases advanced capabilities, including writing Python code to simulate polyethylene melt using GROMACS, repairing topology errors, and generating a JSON protocol for a Suzuki coupling study. The infrastructure supports both NVIDIA CUDA and AMD ROCm, with specialized targets like NVFP4 for B200 and MXFP4 for MI350X.