The Hangzhou startup has begun issuing access keys for its latest research tool, offering the service at no cost to users of compatible third-party clients or its own website. Deep Principle stated that the GPU and server costs for the Hydrogen-1.0 model are funded by the company’s $HYDRO token, allowing the science-focused AI to be available via an OpenAI-compatible base URL.
The open-weight system is built on the GLM-5.3 model from Z.ai and was specialized using internal lab records and discovery trajectories. In performance tests, the model scored 38.3% on the FrontierScience-Research benchmark, surpassing the 33.3% posted by OpenAI’s GPT-6-Astra, and recorded a 93% score on the PhD-level GPQA-Diamonds questionnaire.
