AI-generated concept illustration of a branching chemical reaction network
A reaction map can grow quickly once a model starts proposing what might happen next. Each new branch gives a chemist another possibility to consider, and another claim to check. A plausible-looking pathway may still depend on conditions that the initial question left unspecified.
Deep Principle released Hydrogen-1.0 on November 2. The company adapted GLM-5.3 using data from chemistry, materials science and the life sciences, along with lab records and examples of how research decisions evolved during a project. The release includes downloadable weights, with the Hugging Face model card listing an MIT license.
Where the reaction map branches
In one demonstration, the team used Hydrogen-1.0 to explore an Ireland–Claisen rearrangement network. Its technical announcement describes 27 molecular structures connected by 30 plausible reactions. The case shows how the model and research tools can make competing routes visible in one place.
The map is an inventory of hypotheses. A useful follow-up would identify which branches rely on a narrow set of assumptions, which are robust to changes in the computational setup and which warrant experimental attention. The order matters when a laboratory can afford to investigate only a few candidates.
What the comparison measured
Deep Principle reported stronger results than GPT-6 Astra on its internal scientific computing and forward reaction tasks when comparing their highest tested reasoning settings. Astra led on the team's internal NMR structure task at that setting. These results describe selected evaluations run by the developer. The examples in the announcement also draw on specific tools and research environments.
An outside lab might care more about a smaller question, such as whether the model suggests viable alternatives for a reaction class the lab already knows well. A set of past cases with documented results would help researchers see where the proposals hold up, and where apparently reasonable suggestions lead them astray.
What another lab can reproduce
Downloadable weights give teams a way to run their own cases. The full training datasets and recipes are absent from the weights release, and running the model requires a compatible setup. A reproducible evaluation would therefore need to specify which checkpoint was used, what tools it could call and how candidate reactions were judged.
A short list of predictions with their inputs, assumptions and outcomes would give another group a result it could actually check. For an open-weight model built to assist research, that is a concrete next step from the release itself.
Sources · Deep Principle technical announcement · Hugging Face model card · Release-date reporting (BAAI Hub)