In a seminar on Monday (January 26), GuoXing Aerospace Technology announced it had linked Alibaba's large language model ( LLM ) Qwen3 to its first space computing center, enabling end-to-end reasoning tasks to be performed entirely in orbit.
"This marks the world's first application of a large-scale, versatile AI model from control on Earth to an operational satellite constellation in orbit," said Wang Yabo, executive vice president of the Chengdu-based startup.
In May last year, China launched a new constellation of 12 space computing satellites into orbit, the first cluster of GuoXing Aerospace's space computing project.
In trials, the Qwen3 model completed a variety of experiments, with a number of queries sent from Earth to the satellite, processed directly on the satellite, and the results sent back to the station on Earth.
The whole process takes just two minutes.
As AI drives the increasing need for computing power, a new field is emerging in the technology race as intelligent computing capabilities are brought to space.
In November, a SpaceX rocket launched the Starcloud-1 satellite equipped with an Nvidia GPU into orbit.
Wang outlined his company's ambition to build a massive network of 2,800 dedicated computing satellites by 2035.
The planned constellation includes 2,400 inference satellites and 400 training satellites, which will be placed in Sun-synchronous orbits, dawn-dusk orbits, and low-inclination orbits at altitudes of 500 to 1,000 kilometers.
The constellation is designed to use intersatellite laser links to facilitate high-speed data transfer, with the goal of providing 100,000 petaflops of inference computing capacity and 1 million petaflops of training computing capacity worldwide.
According to Wang, their second and third satellite clusters are expected to be launched this year, with a network of 1,000 satellites completed by 2030.
