Deep Principle has introduced Hydrogen-1.0, a scientific model built by post-training Z.ai’s GLM-5.3 on chemistry, materials science and life-science data, laboratory records and research trajectories. The aim is to help researchers work through experiments and simulations, not just answer scientific questions. The approach is described in the official announcement.
Open weights, with substantial deployment requirements
The release includes MIT-licensed weights. Its model card lists a one-million-token context window and quantized versions for NVIDIA Blackwell and AMD Instinct hardware.
Hugging Face downloads require accepting access conditions. Deployment was tested on eight GPUs, so this is not a lightweight local assistant. The model card also notes that available context and output budgets depend on the inference framework, configuration and runtime resources.
Why it matters for builders
Deep Principle’s demonstrations combine the model with external tools to write and repair simulation code, explore reaction networks and check laboratory protocols in a virtual environment before execution. That makes the release interesting for developers building scientific agents with inspectable, adaptable weights.
The distinction is between an available model and a proven autonomous researcher. The demonstrations and performance comparisons come from the developer, and scientific outputs still need expert verification. Full training datasets and recipes are not included with the weights.
For a different access model in the same field, see our coverage of GPT-6 Rosalind’s restricted research access. Both releases make workflow-level evaluation important: the useful question is whether the system produces correct, reviewable work for the research task at hand.