Deep Principle has released Hydrogen-1.0, a 753B open-weight scientific model designed for chemistry, materials science, and life sciences. The model is built by post-training Z.ai’s GLM-5.3 on domain data, laboratory records, and agent discovery trajectories, and its weights ship under the MIT License. It targets question framing, experiment planning, results analysis, and open-ended discovery as an inspectable alternative to proprietary systems.
The training pipeline combines supervised fine-tuning and reinforcement learning using data from chemistry texts, Deep Principle scientist feedback, electronic lab notebook records, and discovery trajectories. Running on infrastructure using Megatron and SGLang across NVIDIA and AMD hardware, the model supports a 1M-token context and 128K max output. Developers can deploy it on 8 GPUs using BF16 serving, with quantization options including NVFP4 for NVIDIA B200 and MXFP4 for AMD accelerators.
On performance benchmarks, Deep Principle reports that Hydrogen-1.0 scores 38.3% on FrontierScience-Research, outperforming GPT-6-Astra, and roughly matches Kimi-K3 on GPQA-Diamond. In practical demonstrations, the model writes Python scripts for molecular dynamics simulations, analyzes trajectories, grows chemical reaction networks, and compiles JSON protocols for laboratory automation.
Why it matters: Hydrogen-1.0 provides researchers with a powerful, open-weight artificial intelligence assistant for scientific discovery that matches or exceeds proprietary models in specific domain tasks.