Deep Principle Open-Sources Hydrogen-1.0, a 753B Science Model Post-Trained From GLM-5.3 on Lab Records, MIT Licensed

Core Background & Industry Pain Points

Most recent headline AI-for-science results come from closed models whose weights researchers cannot inspect, reproduce or adapt. Deep Principle, whose co-founder and CTO is Chenru Duan, released Hydrogen-1.0 on November 2 as an open-weight scientific model, aiming to show that domain post-training on an open base can catch up with closed flagships.

Architecture Highlights & Internals

It builds on GLM-5.3 and keeps its MoE plus dynamic sparse attention (DSA) design, at about 753B total parameters. Training combines SFT and RL over domain knowledge, scientist feedback, full lab records including failed experiments, and materials discovery trajectories. The infrastructure uses Megatron and SGLang on both CUDA and ROCm.

Authoritative Benchmarks & Measured Scores

All figures are vendor-reported and not yet independently reproduced. It scores 38.3% on FrontierScience-Research vs 33.3% for GPT-6 Astra, and 93.0% on GPQA Diamond. It trails Astra by 3.33 points on NMR structure elucidation. In the experiment-design comparison, Astra and Opus 5.5 were dropped for refusing over half of the requests, so conditions are not fully like-for-like.

Developer Hands-on Guide

Weights are at deep-principle/Hydrogen-1.0 on Hugging Face after accepting the terms; BF16 is about 1.5 TB. The official example serves it with SGLang on 8 GPUs, and generation_config.json must be kept or generation will not stop. NVFP4 and MXFP4 quantized repos lower the hardware bar, and the NeuroSpark API or Mira platform allow trials without hosting. Key claims still need checking against literature or experiment.