Recently, at the stage report meeting of the national key research and development program "Satellite platform technology for cluster/mega constellation applications" project "Cluster/mega constellation collaborative computing architecture and network operating system", the Institute of Computing Technology of the Chinese Academy of Sciences, as the project undertaking unit, introduced its developed aurora space-based large model JigonGPT and on-board intelligent computer load system. As of May 10, 2025, JigonGPT has been in orbit for over 100 days.
A breakthrough has been made in the large model of ground remote sensing. The large basic model of "Kongtian • Lingmou" ground remote sensing proposed by the Aerospace Information Innovation Research Institute of the Chinese Academy of Sciences has developed to the third generation, but limited by the on-board computing resources, the customized autonomous space-based large model and related computing load fields are still blank. Under the funding of the National Key Research and Development Program, researchers Han Yinhe, Liu Yaoqi, and Dr. Zhao Yisheng from the Intelligent Computer Center of the Institute of Computing Technology, together with Beijing Zhiyuan Artificial Intelligence Research Institute and Wuhan University, jointly developed the JigongGPT, a large-scale model of aurora space. This model is based on the Aquila language model of Zhiyuan, integrates visual encoding modules, and constructs a benchmark model with 1.02B parameters. In order to solve the problem of limited computing power and power consumption in space-based computing payloads, the team adopted layered quantization and cross modal distillation techniques. After testing and verification by the China Software Evaluation Center, the model parameters were compressed by 77% while sacrificing only 5.2% of remote sensing interpretation performance (the BLEU-4 index was reduced to 39.91% in bilingual replacement evaluation). Finally, a lightweight version with 0.23B parameters was obtained, successfully breaking through the deployment difficulties in spaceborne computing scenarios and achieving functions such as image parsing, semantic interaction, and instruction generation.
The team has developed a super large file injection software that supports 1024MB file breakpoint continuation, in order to build an injection installation and dynamic update path for large models. Through 12 rounds of intermittent uploads between the ground station and the "Dongfang Huiyan Gaofen01 satellite", overcoming engineering challenges such as intermittent interruptions in the satellite link and channel bandwidth fluctuations, the installation of JigonGPT has been successfully achieved. The large model has been in orbit for over 140 days now.
Dongfang Huiyan Gaofen 01 Star "is the first star of the" Dongfang Huiyan "intelligent remote sensing constellation. The "Oriental Insight" constellation is led by Academician Li Deren of Wuhan University, with Professor Wang Mi serving as the chief engineer. Through the combination of satellite remote sensing and artificial intelligence technology, this constellation provides intelligent analysis and decision-making services for spatiotemporal observations, improving the speed and efficiency of users in obtaining remote sensing information. The onboard intelligent computer payload of the "Dongfang Huiyan Gaofen 01 Star" was developed by a team of researchers led by Han Yin. It uses domestically produced core components and a highly reliable fault-tolerant computing architecture, achieving 100 Tops level intelligent computing capabilities.
Based on JigonGPT, the team has developed the first satellite digital person "Dongfang Xiaoji" with "direct access to social media on satellite", and has established accounts and regularly released them on Weibo and Xiaohongshu. As a multimodal interactive agent, it deeply integrates technologies such as onboard perception, semantic understanding, and network media instruction generation, and can interact with ground users through social media, demonstrating its ability to directly serve end customers.
Space based large-scale models, intelligent agents, and onboard intelligent computers will endow satellites with the ability to perceive, judge, make decisions, and take actions, providing a core technological foundation for the implementation of remote sensing data analysis, intelligence information generation, space-based control and command, and real-time response space application service systems.
