Deep Principle has released Hydrogen-1.0, a model for chemistry, materials science and life sciences. The company says it adapted GLM-5.3, a model from Z.ai, using scientific material, laboratory records and accounts of research agents’ actions. MarkTechPost identifies Hydrogen-1.0 as a 753-billion-parameter model.
The model’s weights—the files needed to run or adapt it—are available on Hugging Face under the MIT License. That makes it possible for researchers and developers to run their own evaluations and build tools around it, rather than relying only on a company-operated service. The release does not include the full training datasets or complete training recipe.
What did Deep Principle test?
Deep Principle reports that Hydrogen-1.0 beat GPT-6-Astra on FrontierScience-Research and improved on its starting model, GLM-5.3, on GPQA-Diamonds, roughly matching Kimi-K3. The company’s announcement gives no scores for those comparisons, so readers cannot judge the size of the reported gains from the text alone.
In the company’s own research-task evaluations, Hydrogen-1.0 scored higher than GPT-6-Astra on scientific computing and forward reaction prediction at all three tested reasoning settings. Against Astra at its maximum setting, the reported leads were 5.13 and 4.76 percentage points. Astra scored 3.33 points higher on NMR structure analysis. Deep Principle also compared the models on 117 experiment-planning questions, but the announcement does not give the models’ scores in its text.
What can researchers run?
Deep Principle demonstrates the model writing and repairing code for a polymer simulation, mapping possible chemical reactions, and assessing candidate structures against an NMR spectrum. Those examples use external tools and research environments; they are demonstrations, not proof that the model’s scientific conclusions are correct.
The Hugging Face listing includes full-precision BF16 weights and lower-precision versions for specific hardware: NVIDIA Blackwell chips or AMD Instinct MI350X and MI355X. The listing says a deployment was tested on eight GPUs, but does not specify the hardware for that test. It also warns that requirements vary by software and setup, and that the model can give incorrect answers. Researchers should check important claims against calculations, literature or experiments.
Why it matters
Researchers can download and adapt Hydrogen-1.0 instead of depending solely on a proprietary science model. But the hardware demands, incomplete training disclosure and company-run evaluations mean they will need to test it on their own work and independently verify its answers.
