Deep Principle releases Hydrogen-1.0 scientific model weights

Deep Principle releases Hydrogen-1.0, a scientific problem solving and research assistance model, as detailed in its Hugging Face model card. The model is based on GLM-5.3 from Zhipu AI and focuses on scientific questions, derivations, experimental planning and open ended research tasks.

The unquantized BF16 weights are available under the MIT License. Separate NVFP4 and MXFP4 versions support NVIDIA B200 and B300 GPUs and AMD Instinct MI350X and MI355X GPUs, respectively. Deep Principle tested BF16 serving with SGLang across eight GPUs.

The 753 billion parameter model reports a win over OpenAI's GPT-6-Astra on the FrontierScience-Research benchmark and, on Deep Principle's own research-task tests, leads GPT-6-Astra by 5.13 points in scientific computing and 4.76 points in forward reaction prediction while trailing by 3.33 points in NMR structure elucidation, MarkTechPost reported.

Training used supervised fine tuning and reinforcement learning with scientific domain knowledge, research scenarios, laboratory records and discovery trajectories. Deep Principle did not release the complete datasets or training recipes.

The model card says outputs can contain incorrect facts, calculations, equations or references. It advises users to verify scientific claims through primary literature, independent calculations or experiments.