AI BriefingDeep Principle releases Hydrogen-1.0, open-weight scientific model outperforming GPT-6 Astra on FrontierScience Research
·2026.11.04 18:00
Key point
Post-trained from GLM-5.3, the model achieves a 38.3% pass rate on FrontierScience Research, surpassing GPT-6 Astra's 33.3% at max thinking.
Details
- Deep Principle released Hydrogen-1.0, an open-weight model post-trained from GLM-5.3 for chemistry, materials science, and life sciences.
- The model achieved a 38.3% pass rate on FrontierScience Research (FSR), leading GPT-6 Astra at max thinking (33.3%) by 5 percentage points.
- Estimated output-side cost on FSR is $0.25 per problem, approximately 72% lower than GPT-6 Astra's $0.88 per problem at max thinking.
- Hydrogen-1.0 scored 92.9% on GPQA Diamond, slightly below Kimi K3's 93.5%, and outperformed GPT-6 Astra on scientific computing (+5.13 points) and single-step forward reaction prediction (+4.76 points).
- Training utilized supervised fine-tuning and reinforcement learning with domain data, scientist feedback, lab records, and discovery trajectories.
- Demonstrated workflows include polymer molecular dynamics using GROMACS, reaction network exploration, NMR interpretation, and machine-readable protocol generation for Suzuki coupling studies.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Ampersand Releases &Tutor, a Tutoring Model Trained on Persimmon Learner SimulationsTraining 1,000 Weights in Gemma 3 270M Unlocks Hidden Sorting and JSON Skills