LOS ANGELES – November 06, 2026 -- Deep Principle released Hydrogen-1.0, an open-weight scientific model that reached a mean best-so-far HOMO–LUMO gap of 4.318 eV after five rounds of transition metal complex design, against 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3, according to the company's own evaluation.
The model was introduced on November 4, 2026. It helps chemistry and materials teams plan experiments, run computational workflows and search for candidates with target properties. Researchers can access it through Deep Principle's AI Scientist, Mira, as well as through model downloads and an API.
Hydrogen-1.0 is post-trained on GLM-5.3 using lab records and failed attempts
Deep Principle built the model by post-training GLM-5.3 on chemistry, materials science and life sciences data, plus records and agent trajectories from its own laboratories. Supervised fine-tuning supplies examples of reasoning, experimental protocols and results analysis. Reinforcement learning applies feedback on research outcomes. Training retains hypotheses, actions, observations and revisions, including failed attempts.
Virtual rehearsal caught a pipetting error in a 12-vial Suzuki coupling protocol
Hydrogen-1.0 converts a brief research goal into an experimental plan and protocols, then interprets results. In a 12-vial Suzuki coupling example, it wrote a JSON protocol and tested it in a virtual rehearsal. The rehearsal flagged an attempt to aspirate without a fresh tip. The model corrected the action order and compiled the revised protocol before execution and analysis handoff. Checks covered station capabilities, dependencies and action prerequisites.
On shared experiment-planning questions, the company reports a higher mean rubric score for Hydrogen-1.0 than for Kimi-K3.
Model repairs a polymer simulation error and maps 30 reactions in a rearrangement network
In a polymer molecular dynamics example, Hydrogen-1.0 wrote Python to build a polyethylene melt and ran GROMACS. It detected and repaired a topology error, restarted the run, and analyzed the corrected 40-nanosecond trajectory for density, molecular packing, diffusion and chain dimensions.
For reaction pathway exploration, the model generated 27 molecular structures through 30 plausible reactions in an Ireland–Claisen rearrangement network. The network covers a five-step sequence of Cope and Claisen rearrangements and silyl shifts. In the selected comparison, the main route had a lower electronic barrier than its displayed alternative at every step.
Trajectory-based training targets fewer attempts to reach desired material properties
Deep Principle trains the model on complete discovery trajectories linking a desired property to candidate selection, synthesis strategies, evaluation and the next decision. In the transition metal complex test, all models started from the same initial population. Each received feedback only from its own recommendations, keeping evaluation histories independent. The company's next research focus is learning across discovery targets and material families, so that one campaign can improve the next.
Benchmarks show gains over GPT-6-Astra, with a deficit on NMR structure elucidation
Deep Principle reports that Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research. It improves on GLM-5.3 on GPQA Diamond and roughly matches Kimi-K3 there. In internal scientific computing and one-step forward reaction prediction tasks, it beat GPT-6-Astra at low, high and maximum thinking settings. Against Astra at maximum thinking, the leads were 5.13 and 4.76 percentage points. Astra at maximum thinking scored 3.33 percentage points higher on NMR structure elucidation.
Four access routes cover Hugging Face, ModelScope, NeuroSpark and Mira
Open-weight downloads are available on Hugging Face and ModelScope. Direct API access runs through NeuroSpark, and the model is also available inside Mira. Deep Principle worked with UniPat and Mind Lab on data, the training pipeline and infrastructure, and with NeuroSpark on training infrastructure.