The Pulse
Deep Principle Releases Hydrogen-1.0 for Scientific Research
Deep Principle released Hydrogen-1.0, an open-weight scientific model post-trained from GLM-5.3 on chemistry, materials science and life-sciences data. The company reports gains on several scientific evaluations, while its model card says t
AI.info Team · Written with AI · fact-checked ·
Deep Principle released Hydrogen-1.0 on Nov. 2, presenting it as an open-weight model for scientific research. The company says it post-trained GLM-5.3 on chemistry, materials science and life-sciences data, adding records and agent trajectories from its own laboratories. The release aims to connect scientific reasoning with experiment planning and materials discovery.
From GLM-5.3 to a research model
Deep Principle’s training approach draws on more than scientific reference material. The company says its electronic laboratory notebook records preserve a research cycle: what a scientist hoped to learn, how an experiment was conducted, what happened and what to investigate next. It says those records help Hydrogen-1.0 plan experiments, write protocols and interpret results.
The model card identifies the released weights as a 753-billion-parameter model based on GLM-5.3. It lists a context length of 1 million tokens and a maximum output length of 128,000 tokens, while cautioning that usable limits depend on the inference framework, service configuration and available resources. The card says the weights use the MIT License.
Benchmark gains come with a task-level exception
Deep Principle reports that Hydrogen-1.0 outperformed GPT-6-Astra on FrontierScience-Research and improved on GLM-5.3 on GPQA-Diamonds, roughly matching Kimi-K3 on those evaluations. The company describes the tests as measures of scientific knowledge and reasoning on challenging questions. It does not provide the benchmark scores in the announcement’s accompanying text.
On three internal research tasks, the results were mixed. Hydrogen-1.0 scored higher than GPT-6-Astra on scientific computing and one-step forward reaction prediction at each of three thinking settings; against Astra at maximum thinking, the reported leads were 5.13 and 4.76 percentage points. Astra scored 3.33 points higher on NMR structure elucidation at maximum thinking. Deep Principle cautions that differences in task difficulty mean the absolute scores should not be read as direct comparisons across tasks.
The examples show what those evaluations cover. In one, the model writes Python to build a polyethylene melt, runs GROMACS, repairs a topology error and analyzes a corrected 40-nanosecond simulation. In another, it maps 27 molecular structures through 30 plausible reactions. A separate laboratory-planning demonstration catches a protocol step that tries to aspirate without a fresh pipette tip, then compiles the corrected instructions in a virtual environment.
Discovery trajectories put choices into the training data
Deep Principle says supervised fine-tuning and reinforcement learning both contribute to training. The first uses worked examples for reasoning, experiment planning, protocol design and results analysis; the second uses feedback on research outcomes. Complete discovery trajectories link a target property to candidate selection, synthesis planning, evaluation and the decision about what to try next.
For one materials-search test, models began with the same population of transition metal complexes and made recommendations over five rounds, each receiving feedback only on its own suggestions. By the fifth round, Hydrogen-1.0 reached a mean best-so-far HOMO–LUMO gap of 4.318 eV, compared with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. Those results come from the company’s reported evaluation, not a claim that the model has independently discovered or synthesized a material.
Weights are available, but the recipe is partial
Deep Principle lists Hugging Face and ModelScope as download routes for the weights, and NeuroSpark and Mira as ways to access the model through an API or the company’s AI Scientist. The Hugging Face model card says users must accept the repository’s access conditions to reach its files. It also says the full training datasets and complete training recipes are not released with the weights.
That leaves a distinction between open weights and a fully reproducible training process. Researchers can obtain the model parameters under the stated license, but the public training account does not include all the data and recipe details needed to recreate the model from scratch.