Deep Principle releases Hydrogen-1.0, an open-weight science model built on GLM-5.3

Deep Principle has published Hydrogen-1.0, a scientific problem solving and research assistance model built on top of GLM-5.3, on Hugging Face. The model is meant to help through text interaction with scientific questions, explaining derivations and organizing responses to open-ended research problems, with the publisher noting that outputs should be checked against the assumptions, constraints and evidence relevant to each task.

Deep Principle's technical blog shows examples of the model paired with external tools in areas like scientific computing, reaction network exploration, NMR structure analysis and experimental planning, though it says these outputs still require scientific review.

The repository holds the unquantized BF16 weights, released under the MIT license, with a context length of 1 million tokens and a maximum output length of 128,000 tokens. Separate quantized versions are also available: NVFP4 weights for NVIDIA Blackwell B200 and B300 hardware, and MXFP4 weights for AMD Instinct MI350X and MI355X chips. Deep Principle says it tested deployment with the SGLang inference framework across 8 GPUs, and that full training datasets and recipes are not released alongside the weights.

Why it matters

An open-weight, MIT-licensed science model gives researchers a model they can run and modify themselves rather than relying on a closed API.

Source: Deep Principle on Hugging Face