Deep Principle Releases Hydrogen-1.0 Open-Weight Scientific Model

Deep Principle released Hydrogen-1.0, an open-weight scientific research model post-trained from GLM-5.3 on chemistry, materials science and life sciences data, laboratory records and discovery trajectories. It scored 38.3% on FrontierScience-Research versus 33.3% for GPT-6-Astra and 93.0% on GPQA-Diamonds, with weights published under an MIT license.
Event version 1.
Key insights
Benchmark performance
Hydrogen-1.0 scored 38.3% on FrontierScience-Research compared with 33.3% for GPT-6-Astra, and 93.0% on GPQA-Diamonds. It roughly matches the open frontier model Kimi-K3 on GPQA-Diamonds.
Internal research scenarios
On internal evaluations, Hydrogen-1.0 scored 76.9% versus 71.8% for GPT-6-Astra on scientific computing, 90.5% versus 85.7% on forward reaction prediction, and 26.7% versus 30.0% on NMR structure elucidation.
Training approach
Hydrogen-1.0 was built by post-training GLM-5.3 on domain data from chemistry, materials science and life sciences, together with laboratory records and agent discovery trajectories. Training used full-cycle lab records covering experiment design, procedure, observations and outcomes.
Open weights and availability
The model weights are published under an MIT license with a 1M-token context length and 128K-token maximum output length. BF16, NVFP4 and MXFP4 versions are available in separate repositories.
Demonstrated research applications
Hydrogen-1.0 built a polyethylene melt, ran GROMACS, diagnosed and repaired a topology error, and analyzed a 40 ns trajectory. In a search for transition metal complexes, it found better candidates with fewer attempts.