"Panstone" Large Model Launches, Empowering a New Research Paradigm

Artificial intelligence is reshaping the fundamental logic of scientific research and accelerating scientific discovery. The integration of AI with science has become a trend, offering unprecedented opportunities to address major scientific and technological challenges. On July 26, the "Panshi·Scientific Foundation Large Model" was officially launched at the 2025 World Artificial Intelligence Conference. Developed by a collaborative team from the Chinese Academy of Sciences, this model is trained using specialized scientific knowledge and data to serve scientific tasks. It demonstrates a deep understanding of various scientific data modalities, including waves, spectra, and fields, and possesses core capabilities such as scientific literature extraction and integration, scientific knowledge representation and reasoning, and scientific tool orchestration and planning. Its goal is to provide robust intelligent support for innovation across all scientific fields. To address the current challenges in "AI + science" research—such as scientific data silos, limited professional reasoning capabilities, and a closed development ecosystem—China's Academy of Sciences joint research team has leveraged its comprehensive natural science discipline system, full-stack AI innovation chain, major scientific facilities, and scientific data resources. This has enabled a systematic approach to drive the transformation of "AI + science" toward a more platform-based and systematic paradigm. The "Panshi·Scientific Foundation Large Model" can manage data and model resources, as well as schedule computational simulation tools, deeply empowering the entire scientific research process—from hypothesis generation to solution planning, simulation, experimentation, and discovery of laws. This marks the emergence of a cross-disciplinary "operating system" for "AI + science." Researchers can now easily call upon the model at every stage of their work, enabling seamless integration of AI into scientific research. The "Panshi·Scientific Foundation Large Model" is built around the core needs of researchers, offering strong and comprehensive scientific professional capabilities to support scientific discovery. Its core architecture uses an heterogeneous hybrid expert model, which is deeply customized from a domestic open-source large model for the scientific domain. It integrates a series of proprietary models tailored for common scientific data modalities and incorporates specialized models like AlphaFold and MatterGen. The model has systematically mastered the core theorems, laws, and knowledge of six major scientific disciplines: mathematics, physics, chemistry, earth sciences, life sciences, and astronomy. It has demonstrated a deep understanding of data modalities such as waves, spectra, and fields. In specific evaluations, the model achieved top performance in international datasets across mathematics, physics, chemistry, materials science, and biology. It also showed leading capabilities in scientific tool invocation and reasoning in authoritative benchmarks such as GAIA and SimpleQA. In the Human-Level Examination (HLE), it performed exceptionally well. The model has been tested on complex reasoning problems and has demonstrated strong scientific capabilities. Based on the "Panshi·Scientific Foundation Large Model," the research team has developed two scientific AI agents: "Panshi·Literature Compass" and "Panshi·Tool Scheduler." The "Literature Compass" assists researchers in reading articles, writing reviews, and evaluating research topics and technical paths. It has accessed 170 million scientific papers and real-time open-source scientific information, enabling it to deeply understand scientific data that includes formulas and charts. It can thoroughly analyze thousands of papers in one go, reducing the time previously required for literature reviews from 3 to 5 days to just 20 minutes. The "Tool Scheduler" lowers the barriers to using scientific tools by autonomously planning and invoking over 300 scientific computing tools, enabling efficient orchestration and convenient access. It can automatically identify research tasks, intelligently arrange and schedule the optimal tool chain, improving research efficiency and allowing users to quickly build custom research applications by integrating their own AI agents and tools. The model is already driving transformation and accelerating scientific discovery across multiple fields. In life sciences, it has been used to develop the X-Cell Digital Cell Large Model, which enables end-to-end modeling from gene sequences and the central dogma to cell phenotypes, supporting the automation of the entire process of target discovery. X-Cell is being used for regulatory network analysis, virtual cell experiments, and target discovery. For instance, in the case of identifying protein interaction-based drug targets, the research efficiency has increased by more than ten times compared to existing methods. In high-energy physics, researchers at the Beijing Electron-Positron Collider are using the model to automatically decompose and plan particle physics research tasks, generating analysis programs that cover all stages of the particle physics workflow. This has significantly improved the speed of particle simulations and reconstruction efficiency, helping to explore the fundamental composition of matter and the basic laws of the universe. In mechanics, the model’s strong ability to understand and predict scientific data has enabled efficient calculations of surface pressure fields for high-speed train models in various fluid environments, supporting the design of high-speed train configurations. It is also helping scientists improve experimental efficiency in chemical synthesis, achieve more accurate molecular structure predictions, and enable intelligent global resource scheduling and analysis in astronomical observations. As it continues to address real scientific needs, the "Panshi·Scientific Foundation Large Model" is constantly evolving to enhance its practicality and reliability. The model is now becoming a versatile expert in scientific knowledge, a super analyst of literature, and a precision calculator for complex scientific problems. It is gradually evolving into a strategic advisor and intelligence support for scientific research. To promote collaborative innovation in "AI + science," the Institute of Automation at the Chinese Academy of Sciences has launched the "Scientific Foundation Large Model Ecosystem Alliance" with over 40 research institutes, universities, and enterprise partners. The initiative is also working on adapting to domestic computing power and building an open, self-controlled "AI + science" ecosystem. It aims to contribute a Chinese solution to the global academic community and unlock new possibilities for scientific research. The "Panshi·Scientific Foundation Large Model" is now fully open-source, and users are encouraged to experience it. Visit the official website: https://scienceone.ia.ac.cn/.