CAS Unveils Two "Kunyuan" Large Models

On July 29, during the third Coastal Zone Conference held in Yantai, Shandong Province, Dr. Su Fen zhen, a researcher at the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, and chief scientist of the "Kunyuan" large models, unveiled two independently developed artificial intelligence models: "Kunyuan·Gan Dongnan" and "Kunyuan·Ni Qian Hai." Designed to address land and marine challenges respectively, these models establish an integrated intelligent monitoring and simulation system for terrestrial and marine environments. The launch marks a significant milestone in the fusion of remote sensing, oceanography, and artificial intelligence, with potential to support precise, rapid regional remote sensing detection and global ocean modeling. The "Kunyuan·Gan Dongnan" model focuses on land resource and environmental monitoring in Southeast Asia. Leveraging self-developed sample generation techniques and efficient parameter fine-tuning methods, it achieves high-precision land cover mapping across Southeast Asia from 1990 to 2020. The model demonstrates an overall accuracy exceeding 92% across seven major land cover categories. Notably, it operates with remarkable efficiency—processing the entire region using only consumer-grade NVIDIA GeForce RTX 4090 graphics cards in just 3 to 4 hours. This breakthrough redefines traditional resource and environmental monitoring approaches, ushering in a new era of intelligent, high-efficiency analysis for the region. The "Kunyuan·Ni Qian Hai" model targets the challenge of limited data and theoretical constraints in deep ocean (kilometer-level) oceanographic research. Scientists developed a physics-guided spatiotemporal intelligent modeling framework that incorporates time-space correlation information into an innovative autoregressive pre-training architecture. By applying a progressive fine-tuning strategy and assimilating over 1.3 million Argo trajectory data points, the model achieves high-precision spatiotemporal reconstruction of deep ocean currents. Its accuracy surpasses traditional methods by up to 25%, enabling deeper insights into mid-level ocean current dynamics, material transport mechanisms, and multiscale interactions. This advancement enhances scientific understanding of oceanic processes and their role in global climate systems. Looking ahead, the research team plans to integrate vast amounts of geospatial data to expand monitoring capabilities, reduce technical barriers, and support sustainable development across land and sea. Key priorities include improving the detection and forecasting of complex coastal zone phenomena such as land cover evolution and marine disaster responses. The ultimate goal is to build a powerful, scalable technological platform that advances global research and applications in coastal zone science.