A model for chemistry and materials teams
Deep Principle introduced Hydrogen-1.0 on November 4, 2026, an open-weight scientific model the company said is designed to help chemistry and materials research teams plan experiments, run computational workflows and explore candidates for desired material properties. Researchers can reach it through Mira, Deep Principle's AI Scientist, through model downloads and through an API, the company said.
Trained on laboratory records
Hydrogen-1.0 was built by post-training GLM-5.3 on chemistry, materials science and life sciences data, together with records and agent trajectories from Deep Principle's own laboratories. Supervised fine-tuning supplies examples of reasoning, experimental protocols and analysis of results, while reinforcement learning uses feedback on research outcomes. The training keeps hypotheses, actions, observations and revisions, including attempts that failed.
4.318 eV after five rounds
In a transition metal complex design evaluation, the model proposed new complexes over five rounds while seeking a larger HOMO-LUMO gap. All models started from the same population and each saw only the results of its own recommendations; by round five, the mean best-so-far gap reached 4.318 eV for Hydrogen-1.0, against 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. The company also reported a higher mean rubric score than Kimi-K3 on shared experiment-planning questions, a lead over GPT-6-Astra on FrontierScience-Research, an improvement on GLM-5.3 on GPQA Diamond, where it roughly matched Kimi-K3, and leads of 5.13 and 4.76 percentage points over GPT-6-Astra at its maximum thinking setting on internal scientific computing and one-step forward reaction prediction tasks. GPT-6-Astra at that setting scored 3.33 percentage points higher on NMR structure elucidation.
Where to get it
Deep Principle lists four routes to the model: open-weight downloads on Hugging Face and ModelScope, direct API access through NeuroSpark, and use inside Mira. The company said it worked with UniPat and Mind Lab on data, the training pipeline and infrastructure, and with NeuroSpark on training infrastructure, and that it welcomes collaboration with the open-model community.
