What was announced?

Deep Principle has released Hydrogen-1.0, an open-weight model built for scientific research. It is a post-trained derivative of Z.ai's GLM-5.3, adapted with chemistry, materials science and life sciences data. The weights are distributed under the MIT License.

Technical specs

  • 753-billion-parameter Mixture-of-Experts architecture, with roughly 40 billion active parameters.
  • 1-million-token context window and up to 128K output tokens.
  • Tested on 8 GPUs with SGLang; supports NVFP4 for NVIDIA B200/B300 and MXFP4 for AMD MI350X/MI355X.

How it was trained

The pipeline combines supervised fine-tuning with reinforcement learning. Data comes from four sources: domain knowledge, task-specific feedback from Deep Principle scientists, full-cycle electronic lab notebook records captured through Mira, and discovery trajectories that include failed attempts. Infrastructure runs on Megatron and SGLang across NVIDIA CUDA and AMD ROCm.

Results

According to Deep Principle's own reporting, the model scores 38.3% on FrontierScience-Research versus 33.3% for GPT-6-Astra. It reports 93.0% on GPQA-Diamond, roughly matching Kimi-K3. It leads on scientific computing and forward reaction prediction but trails GPT-6-Astra by 3.33 points on NMR structure elucidation. Full datasets and training recipes were not released.