Deep Principle launches open-weight Hydrogen-1.0 for scientific research

A model trained on scientific workflows

Deep Principle introduced Hydrogen-1.0 on November 4, 2026. The company says it was built by post-training GLM-5.3 on chemistry, materials science and life sciences data, along with records and agent trajectories from its laboratories.

The training process combines supervised fine-tuning, which provides examples of reasoning, experimental protocols and results analysis, with reinforcement learning based on research outcomes. According to Deep Principle, the training preserves hypotheses, actions, observations and revisions, including failed attempts. The approach is intended to give the model experience with how research decisions evolve rather than only with successful results.

Hydrogen-1.0 also learns from Deep Principle’s electronic laboratory notebook records. Those records, supported by documentation from Mira, connect experimental intent with procedures, observations and outcomes.

From experiment plans to simulations

For AI-assisted experiment design, Hydrogen-1.0 can turn a brief research goal into an experimental plan, produce protocols and interpret results. In a 12-vial Suzuki coupling example, the model created a JSON protocol and tested it in a virtual rehearsal. The rehearsal identified an attempt to aspirate without a fresh tip, after which Hydrogen-1.0 changed the action order and successfully compiled the revised protocol. The checks covered station capabilities, dependencies and action prerequisites.

In another example, the model translated a polymer materials question into a simulation workflow. It wrote Python to build a polyethylene melt, ran GROMACS, detected and repaired a topology error, then restarted the simulation. It examined the corrected 40-nanosecond trajectory for density, molecular packing, diffusion and chain dimensions.

For reaction pathway exploration, Hydrogen-1.0 generated 27 molecular structures through 30 plausible reactions in an Ireland-Claisen rearrangement network. Deep Principle says mapping these routes can help researchers compare pathways, identify bottlenecks and prioritize further investigation.

Discovery evaluations and access

Deep Principle evaluated Hydrogen-1.0 as a system that repeatedly selects and evaluates material candidates. In a transition metal complex design test, the model recommended new complexes over five rounds while seeking a larger HOMO-LUMO gap. By the fifth round, the mean best-so-far gap was 4.318 eV for Hydrogen-1.0, compared with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. Each model began with the same initial population and received only feedback from its own recommendations.

The company also reports that Hydrogen-1.0 outperformed GPT-6-Astra on FrontierScience-Research and improved on GLM-5.3 on GPQA Diamond, where it roughly matched Kimi-K3. In internal scientific computing and one-step forward reaction prediction tasks, it scored higher than GPT-6-Astra at its low, high and maximum thinking settings. GPT-6-Astra at maximum thinking scored higher on NMR structure elucidation.

Researchers can access Hydrogen-1.0 through open-weight downloads on Hugging Face and ModelScope, direct API access on NeuroSpark, or Mira. Deep Principle says its next research focus is extending learning across discovery targets and material families, allowing experience from one campaign to inform later work.