Deep Principle introduces Hydrogen-1.0 for scientific research Deep Principle introduces Hydrogen-1.0 for scientific research

Learning from experiments

The company describes supervised fine-tuning and reinforcement learning using scientific knowledge, laboratory records and sequences of research decisions. Those sequences preserve what a researcher tried, what happened and how the next choice changed. One demonstration writes and repairs a molecular-dynamics workflow before examining a 40-nanosecond simulation. Another rehearses a laboratory protocol virtually, catches an ordering mistake and corrects it within the virtual rehearsal. The announcement lists downloads through Hugging Face and ModelScope, API access through NeuroSpark, and integration into its Mira research assistant.

Weights are not the complete training recipe

The unquantized model card lists 753 billion parameters and MIT-licensed weights. Hugging Face requires users to accept access conditions and share account information before downloading that repository. The card lists a million-token context and a 128,000-token output ceiling, while making their usable size dependent on the serving configuration and resources. Deep Principle reports testing the unquantized model with SGLang across eight GPUs. Its client example connects to an existing inference service rather than loading the model. Independently authored code and documentation carry Apache 2.0 terms; third-party files retain their own licenses.

The smaller format still needs substantial hardware

The separate NVIDIA quantized checkpoint has a more specific deployment description: eight B300 GPUs running SGLang. It uses four-bit routed experts, eight-bit dense feed-forward layers near the beginning, and retains higher precision for other components, including attention and shared experts. The supplied server example limits the deployed context to 262,144 tokens, below the model card’s headline capacity. Input and generated text must fit that configured window together. Its cache also uses eight-bit precision. The documentation keeps the tokenizer, chat template, generation settings, quantization files and weight index together so serving uses matching stopping rules.

Research questions and research projects differ

Some internal evaluation questions were adapted from the team’s earlier Scientific Discovery Evaluation work. That paper, revised May 8, separates question-level accuracy from project outcomes. Experts built scenarios around real research projects, reviewed questions for correctness and assembled 1,125 items across biology, chemistry, materials and physics. Project tests instead require hypotheses, simulation or experimental feedback, and subsequent revision. The researchers found that scenario scores did not always predict project performance, with different models leading different projects. This distinction explains why a science quiz and a completed investigation are separate tests: identifying the right answer once does not measure how a system responds when an experiment changes the available evidence.