Deep Principle Introduces Hydrogen-1.0 for AI-Driven Materials Discovery and Chemistry Research

LOS ANGELES, Nov. 6, 2026 — On November 4, 2026, Deep Principle introduced Hydrogen-1.0, an open-weight scientific model that helps chemistry and materials research teams plan experiments, run computational workflows and explore candidates for desired material properties. Researchers can use Hydrogen-1.0 through Mira, Deep Principle's AI Scientist, as well as model downloads and an API.

Built by post-training GLM-5.3, Hydrogen-1.0 draws on chemistry, materials science and life sciences data, together with records and agent trajectories from Deep Principle's own laboratories. Supervised fine-tuning provides examples of reasoning, experimental protocols and results analysis; reinforcement learning uses feedback on research outcomes to improve decisions. Training preserves hypotheses, actions, observations and revisions, including failed attempts.

From chemistry research goals to laboratory protocols

For AI-assisted experiment design, Hydrogen-1.0 turns a brief research goal into an experimental plan, generates precise protocols and interprets results. It learns from Deep Principle's electronic laboratory notebook records, supported by Mira's agentic documentation, which connect experimental intent, procedures, observations and outcomes. In the company's evaluation across shared experiment-planning questions, Hydrogen-1.0 achieved a higher mean rubric score than Kimi-K3.

A laboratory compilation example shows how Hydrogen-1.0 prepares instructions for particular equipment. In a 12-vial Suzuki coupling example, Hydrogen-1.0 wrote a JSON protocol and tested it in a virtual rehearsal. The rehearsal detected an attempt to aspirate without a fresh tip. Hydrogen-1.0 corrected the action order and successfully compiled the revised protocol before final execution and analysis handoff. The checks covered station capabilities, dependencies and action prerequisites.

Running simulations and exploring reaction pathways

In a polymer molecular dynamics example, Hydrogen-1.0 translated a materials question into an executable simulation workflow. It wrote Python to build a polyethylene melt, ran GROMACS, detected and repaired a topology error, and restarted the simulation. It then examined the corrected 40-nanosecond trajectory, including 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. The network includes competing branches and return paths around 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 each step. Mapping these possibilities helps researchers compare routes, identify potential bottlenecks and prioritize further investigation.

Selecting materials through repeated evaluation

Hydrogen-1.0 approaches materials discovery as a sequence of candidate choices and evaluations. Deep Principle trains the model on complete discovery trajectories that connect a desired property to candidate selection, synthesis strategies, evaluation and the next research decision. This gives the model experience in choosing what to investigate and using the resulting evidence to guide further exploration.

In a transition metal complex design evaluation, Hydrogen-1.0 recommended new complexes over five rounds, seeking a larger HOMO–LUMO gap. All models began with the same initial population. Each received only the results of its own recommendations as feedback for the next round, maintaining independent evaluation histories. By round five, the mean best-so-far gap reached 4.318 eV for Hydrogen-1.0, compared with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3.

The result points to the potential of learning from discovery trajectories to help researchers find materials with desired properties in fewer attempts. Deep Principle's next research focus is to extend learning across discovery targets and material families, so that experience from one campaign can improve the model's approach to the next.

Scientific evaluations and access

In Deep Principle's evaluations, Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research and improves on GLM-5.3 on GPQA Diamond, roughly matching Kimi-K3 on the latter benchmark. In internal scientific computing and one-step forward reaction prediction tasks, Hydrogen-1.0 scored higher than GPT-6-Astra running at each of its low, high and maximum thinking settings. Against Astra at maximum thinking, its leads were 5.13 and 4.76 percentage points, respectively. Astra at maximum thinking scored 3.33 percentage points higher on NMR structure elucidation.

Deep Principle offers four routes to Hydrogen-1.0: open-weight downloads on Hugging Face and ModelScope, direct API access on NeuroSpark, and use within Mira.

Deep Principle worked with UniPat and Mind Lab on data, the training pipeline and infrastructure, and with NeuroSpark on training infrastructure. The company welcomes collaboration with the open-model community to develop scientific models and advance scientific discovery.

SOURCE Deep Principle