Deep Principle released Hydrogen-1.0, a research model trained on Zhipu’s GLM-5.3, according to a Beating AI News Flash report carried by BlockBeats. The report said Hydrogen-1.0 scored 38.3 in OpenAI’s FrontierScience-Research evaluation, ahead of GPT-6 Astra at 33.3 and the unmodified GLM-5.3 at 24.4.
Training and demonstrations
Deep Principle said it trained the model on complete laboratory records, including failed experiments and later revisions, rather than only on published results. In demonstrations described by the report, Hydrogen-1.0 wrote code to simulate plastic-molecule motion, corrected a faulty simulation setting, and designed procedures for 12 chemical reactions. It also caught a missing pipette-tip change in a virtual lab workflow.
Open weights, private data
The team ran five rounds of new-material screening and said its final candidate exceeded GPT-6 Astra and Kimi-K3 on the team’s stated metrics. The model weights were released under the MIT license, while the laboratory data used for training was not made public. The benchmark and demonstration results are reported by BlockBeats and are not an independent audit or a forecast.
Source: BlockBeats English report, published Nov. 3, 2026.