Deep Principle Releases Hydrogen-1.0 Open-Weight Science Model
Deep Principle released Hydrogen-1.0, an open-weight science model post-trained on GLM-5.3 using full lab notebook records and discovery trajectories, including failed attempts. It scores 38.3% on FrontierScience-Research, surpassing GPT-6-Astra's 33.3%, though it trails in NMR structure elucidation. The 753B parameter model features a 1M context window, is licensed under MIT, and supports BF16/NVFP4/MXFP4 quantization.
Full Report
▲ Marktechpost AI · 2026-11-02 18
▲ Deep Principle just shipped Hydrogen-1.0, an open-weight science model that learns from how experiments actually unfold, not just from papers.
The key idea is post-training GLM-5.3 on full lab notebook records and discovery trajectories. The model sees intent, procedure, observations and outcomes together, including failed attempts. RL then trains it on complete discovery campaigns.
It scores 38.3% on FrontierScience-Research vs 33.3% for GPT-6-Astra, and 93.0% on GPQA-Diamond (self-reported). In 5-round transition metal complex design, it reached a 4.318 eV gap vs 3.689 eV for GPT-6-Astra. It is a trade-off though: GPT-6-Astra still leads by 3.33 points on NMR structure elucidation, and most comparisons are internal evals.
753B parameters, 1M context, MIT license. BF16, NVFP4 (B200/B300) and MXFP4 (MI350X/MI355X) weights, plus a hosted API via NeuroSpark......
Full analysis:
