753B Parameters, MIT License: Hydrogen-1.0 Makes Its Scientific Model Weights Fully Public

Deep Principle has released Hydrogen-1.0, an open-weight model focused on chemistry, materials science and life science. As covered by MarkTechPost contributor Asif Razzaq, this model is post-trained on Z.ai's GLM-5.3, with training materials including domain-specific datasets, lab records, and agent discovery trajectories. Its weights are released under the MIT License, and the original report confirms the model reaches a parameter scale of 753B.

For developers, the most direct change brought by open weights is full control. Model parameters are available for download, so institutions can deploy, fine-tune and audit the model in their own private environments, instead of relying solely on closed-source API calls. In scientific scenarios where lab records, unpublished data and internal research workflows are highly sensitive, this level of controllability delivers tangible practical value. The MIT License also lowers legal barriers for usage: at least per the license terms, it leaves extensive room for commercial use and custom modification.

Another key detail of Hydrogen-1.0 is its benchmark performance. The report states it has outperformed GPT-6-Astra on the FrontierScience-Research benchmark. This result comes from materials published by the model's developer, which proves the team has made targeted optimizations for scientific reasoning and discovery workflows, but independent reproduction is still required to validate the result. Evaluating scientific models is inherently complex: task design, data contamination risks and review standards all affect the final conclusion. When seeing this performance win, it is important to note that it is a reported test result, not a universally verified final conclusion.

The agent discovery trajectories in the training data point to a clear design direction: the model does not only learn static domain knowledge, but also absorbs action trajectories from real scientific discovery processes. For developers working on AI for Science, this is far more valuable than simply chasing high benchmark scores. It means the model can be potentially used to assist with experiment planning, hypothesis generation, tool calling and result interpretation. Of course, to deploy these capabilities in real labs, the model still needs to pass strict checks for reliability, interpretability and operation cost.

The 753B parameter scale also sends a clear message: open weights do not mean the model can be run by every user. Ordinary users can hardly deploy a model of this size on local devices, and the parties that can actually make use of it are mostly labs, enterprises and research institutions with sufficient computing resources. The MIT License solves the rights issue, but high computing costs still exist. For most developers, a more realistic path may be to research its fine-tuning methods and evaluation performance, or wait for the community to release smaller distilled versions.

The most noteworthy part of Hydrogen-1.0 is that it pushes open-weight scientific models one step further in real-world adoption. Whether it can become a mainstream research infrastructure will depend on subsequent independent evaluations, real lab experiment performance and the growth of its surrounding community ecosystem. The original article link is here: Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery.