Deep Principle Releases Hydrogen-1.0 Research Model
The open-weight model posted a 38.3% pass rate on OpenAI's FrontierScience Research evaluation, ahead of GPT-6 Astra's 33.3%.
Updated on Nov. 3, 2026 in Chemistry
Deep Principle has released Hydrogen-1.0, a research model built on GLM-5.3, with its weights available under the MIT license. The company reported that the model scored 38.3% on OpenAI's FrontierScience Research evaluation, compared with 33.3% for GPT-6 Astra.
Why it matters
The release makes a model trained on detailed laboratory records available to researchers, offering a look at how AI systems might support experimental science. The results are reported by Deep Principle rather than independently verified, and the full training datasets and recipes were not released with the weights.
The evidence is a developer-reported pass rate on OpenAI's FrontierScience Research evaluation, with Hydrogen-1.0 at 38.3% against 33.3% for GPT-6 Astra. Deep Principle's figures are based on pass rates rather than raw point scores, and the results have not been independently verified.
The players
Deep Principle
Deep Principle is the developer of Hydrogen-1.0, which it built on the GLM-5.3 base model and released under the MIT license.
OpenAI
OpenAI is the artificial intelligence company that created the FrontierScience Research evaluation on which the models were tested.
The details
Deep Principle trained Hydrogen-1.0 on laboratory records describing experiment design, errors, and changes made during research, a type of material that scientific papers typically omit. In a polyethylene melt simulation, the model detected and repaired a topology error and restarted the simulation, and in a virtual lab rehearsal it prepared a JSON protocol for a 12-vial Suzuki coupling study. Deep Principle also ran five rounds of new-material screening, with each round informed by the previous evaluation.
Timeline
Article on Hydrogen-1.0 results and release published November 3, 2026.
The Big Picture
Hydrogen-1.0's release under the permissive MIT license places it among open-weight models that researchers can inspect and adapt directly. Its reported results on OpenAI's FrontierScience Research evaluation put an openly licensed model ahead of GPT-6 Astra, a closed comparison point in this benchmark.
Researchers who want to experiment with Hydrogen-1.0 can download its weights under the MIT license now. The full training datasets and recipes were not released, which limits how closely others can reproduce the model's training.
The takeaway
Readers following AI in science can watch whether independent evaluations confirm Deep Principle's reported pass rates. The release suggests that training on records of failed and revised experiments may be a route to more useful research assistants.
Further reading
For related coverage of experimental methods, see the Chemistry section.
Source note: This article includes information reported by TokenPost.
