Hydrogen-1.0 brings a 753B open-weight model to scientific questions

Deep Principle has published Hydrogen-1.0, a 753-billion-parameter scientific research assistant built by post-training GLM-5.3. The public model card describes unquantized BF16 weights under the MIT license, alongside separately published quantized versions. The original repository was created October 31 and updated November 2; this article was researched November 9.

A model for scientific workflows

The company’s technical overview describes training with chemistry, materials-science and life-science data, laboratory records and task trajectories. It presents examples involving scientific computing, reaction networks, structure analysis and experiment planning. Those examples combine a model with particular tools and environments; they are not evidence that every deployment can reproduce the outcomes.

Deep Principle reports results on research benchmarks, including comparisons with larger hosted systems. Those are publisher evaluations, not independent measurements. The card warns that a model may invent facts, citations, calculations or explanations and that scientific conclusions need verification against literature and experiments.

Open weights still require hardware

The BF16 checkpoint is exceptionally large. The model card’s example deployment uses SGLang across eight GPUs; the listed 1 million-token context and 128,000-token maximum output are capabilities described by the publisher, not guarantees at practical runtime settings. Available context depends on memory, implementation and configuration.

The weights are public, but the card says full training datasets and complete recipes are not included. Our view: making a scientific model inspectable is valuable, yet the practical test is whether scientists can trace an answer to evidence and independently reproduce the result. A plausible explanation is a starting point, not a discovery.