Hydrogen-1.0 brings open weights to scientific AI

Deep Principle has released Hydrogen-1.0, a scientific AI model whose downloadable weights let outside researchers run their own tests. Its ModelScope listing is dated November 2, 2026. The release provides access to the trained model, while leaving important parts of the process that produced it undisclosed.

The company says it adapted GLM-5.3 using chemistry, materials-science and life-science data, laboratory records and sequences of decisions made during discovery work. Its release announcement describes scientific computing, reaction prediction and structure analysis as test cases.

The Hugging Face model card lists BF16 weights under the MIT license, with separate NVFP4 and MXFP4 versions targeting NVIDIA Blackwell and AMD Instinct hardware. Access to the Hugging Face files requires accepting conditions to share account information. The card reports testing with SGLang on eight GPUs; downloading weights is not the same as having a hosted service or a model ready for an ordinary laptop.

The comparison changes with the task

Deep Principle’s internal evaluations put Hydrogen ahead of GPT-6 Astra on scientific computing and forward reaction prediction, but behind it on NMR structure elucidation. The figures below compare against Astra at maximum thinking. They are developer-reported results, rather than independent measurements.

A separate experiment-planning comparison scored Hydrogen and Kimi K3 on 117 shared questions. Deep Principle excluded GPT-6 Astra and Claude Opus 5.5 because they rejected more than half the requests. That restricted comparison does not establish a general advantage over all four models. The release’s demonstrations also combine model outputs with external tools and research environments.

Testing a model and rebuilding it are different tasks

The model card states that complete training datasets and recipes are not included. This limits attempts to reproduce the training process even when the resulting weights can be examined or adapted. It also warns that answers may contain false facts, faulty calculations or references, and that quantized versions can behave differently from the BF16 release.

That distinction goes beyond licensing. The Open Source Initiative’s AI definition calls for model parameters, training and inference code, and enough information about training data for a skilled person to build a substantially equivalent system. It does not require publishing every unshareable dataset, but does require detailed descriptions of such data. A weights release alone therefore does not establish full training reproducibility.

For a laboratory deciding whether to use Hydrogen, the next useful evidence would be task-specific tests of the actual version it can deploy: whether calculations check out, cited work exists, and proposed procedures fit the equipment and constraints. Access to weights makes those tests possible outside the developer’s service; it does not supply their results.