Deep Principle, an AI-for-materials startup based in Hangzhou, has released Hydrogen-1.0, an open-weight model built for research in chemistry, materials science and life sciences. The model is available under the MIT license and, according to the company, beats GPT-6-Astra on a scientific research test, while trailing it on a structure analysis task.
Hydrogen-1.0 was not built from scratch. Deep Principle took GLM-5.3, the open model from Z.ai, and post-trained it using supervised fine-tuning combined with reinforcement learning. The training data comes from four sources: domain knowledge in chemistry, materials science and life sciences; research scenarios with feedback from the company's scientists; laboratory records captured through an electronic lab notebook; and discovery trajectories, meaning sequences of choices and results, including failed attempts.
Image: digg.com — https://digg.com/science/ftqpdrb6
On the numbers, Deep Principle reports 38.3% on FrontierScience-Research against 33.3% for GPT-6-Astra, and 93.0% on GPQA-Diamond. These are self-reported figures. In internal tests, Hydrogen-1.0 leads GPT-6-Astra by 5.13 percentage points on scientific computing and by 4.76 points on forward reaction prediction, while GPT-6-Astra leads by 3.33 points on NMR structure elucidation.
The model can write Python to build a polyethylene melt, run GROMACS, detect and fix a topology error, and analyze a 40-nanosecond trajectory. It can also grow a reaction network of 27 structures and rank candidate products from an NMR trace.
Weights are available on Hugging Face and ModelScope, with NVFP4 builds for NVIDIA B200 and B300 and MXFP4 builds for AMD MI350X and MI355X. The model has a 1-million-token context window and up to 128,000 output tokens, with 753 billion total parameters and about 40 billion active per token. Full training datasets and recipes are not released. API pricing has not been disclosed.
No comments yet. Be the first!