TL;DR
- Size: 753B parameters (mixture-of-experts, approximately 40B active per the GLM-5.3 base), with a 1M-token context window and 128K maximum output.
- Runs on: Eight GPUs with SGLang (tested); NVFP4 quantization for NVIDIA B200/B300 and MXFP4 for AMD MI350X/MI355X.
- Performance: Outperforms GPT-6-Astra on FrontierScience-Research and roughly matches Kimi-K3 on GPQA-Diamond.
- Best result: 38.3% on FrontierScience-Research versus 33.3% for GPT-6-Astra (self-reported).
- Weakest result: Trails GPT-6-Astra (max thinking) by 3.33 points on NMR structure elucidation.
What Is Hydrogen-1.0?
Hydrogen-1.0 is a scientific research assistant model from Deep Principle, a Hangzhou-based AI-for-materials company. According to its model card, it is a derivative of Z.ai's GLM-5.3, post-trained on domain data, laboratory records, and agent discovery trajectories.
Deep Principle has released Hydrogen-1.0 as an open-weight scientific model for chemistry, materials science, and the life sciences. For developers, the release offers openly available weights that report a FrontierScience-Research win over GPT-6-Astra. The weights ship under the MIT License.
Positioning in the 2026 Scientific AI Landscape
The release lands amid intensifying competition in domain-specific scientific reasoning models. As of late 2026, open-weight releases in the 700B-plus parameter class remain rare, and most frontier scientific benchmarks are still dominated by proprietary systems. Hydrogen-1.0's MIT-licensed weights—combined with day-one support for SGLang and low-precision formats such as NVFP4 and MXFP4—position it as one of the more deployment-friendly options for research teams running their own inference infrastructure rather than relying on hosted APIs.
via MarkTechPost