CAS Unveils “Panshi 100” Model System to Reshape AI for Science Innovation

As the new round of technological revolution advances at an accelerated pace, AI for Science (Artificial Intelligence for Science)—a deep integration of machine learning and scientific research—is profoundly reshaping the global paradigm of scientific and technological innovation. On 28 April, the Chinese Academy of Sciences (CAS) officially launched the “Panshi 100” model system in Beijing, pooling the academy’s strengths to leverage interdisciplinary collaboration, systematic layout and institutionalized research, thus shifting AI for Science R&D from a “decentralized and closed workshop model” to a “collaborative and open platform model”.

The newly released “Panshi 100” system is built on the “Panshi·Scientific Foundation Large Model” as its cornerstone, supported by a cluster of discipline-specific large models and supplemented by application models and agents for segmented scientific research scenarios, forming a comprehensive and efficient digital intelligent innovation platform. Developed under the leadership of the Institute of Automation of CAS, in conjunction with the Computer Network Information Center, National Science Library, over a dozen field research institutes and two industrialization platforms of the Institute of Automation—Zhongke Wenge and Zhongke Zidong Taichu—the “Panshi·Scientific Foundation Large Model” serves as an intelligent base trained on professional scientific corpora and data to support scientific tasks.

Its upgraded 1.5pro version, equipped with three scientific modal models—wave base, spectrum base and field base—has achieved leapfrog improvements in scientific knowledge reasoning, multi-modal understanding and generation, and model reliability, based on 6.5 million self-constructed high-quality scientific reasoning data. It has reached flagship model levels in scientific knowledge Q&A and agent long-range reasoning rankings, and achieved the current optimal performance in multiple authoritative evaluations related to scientific image understanding and operation.

The model system offers three core functions—Literature Compass, Innovation Evaluation and Agent Factory—to empower the entire scientific research process. The 1.5pro version cuts the cycle of in-depth research surveys by more than half and improves the efficiency of producing papers, PPTs and reports by 5 to 10 times. It also boasts over 2,000 accumulated scientific research tools, supporting more than 10 segmented scientific research fields.

Building on this foundation, CAS has refined AI-Ready scientific issues and built a cluster of discipline-specific large models covering mathematics, physics, materials, astronomy, aerospace, geoscience, biology and environmental science. For instance, “Panshi·Yuhang” in aerospace science is the world’s first aerospace discipline-specific model with in-depth domain cognition, while “Panshi·Shuzi Xibao” in life science discovered three previously unknown drug targets in 30 days, all verified by wet experiments.

It is understood that the “Panshi 100” model system has been promoted and applied in more than 50 CAS units, covering over 100 scientific research scenarios. It has demonstrated great potential in typical scenarios such as high-speed rail flow field reconstruction, spectral recognition, material discovery, astronomical observation, Qinghai-Tibet scientific expeditions and marine forecasting, contributing to China’s progress in AI-driven scientific innovation and offering a Chinese solution for building a global academic community.