BEIJING -- The Chinese Academy of Sciences (CAS) unveiled the ScienceOne model on Saturday at the 2025 World Artificial Intelligence Conference, marking a significant advance in AI-driven research innovation.

Designed to comprehend complex scientific data, including waveforms, spectra, and fields, ScienceOne integrates capabilities for literature extraction, knowledge reasoning, and tool orchestration, aiming to serve as an intelligent foundation for technological innovation across various fields.

AI is transforming fundamental research methodologies and accelerating scientific discovery, providing unprecedented opportunities to address humanity's most pressing scientific challenges. Yet current approaches predominantly rely on fine-tuning general-purpose models with domain-specific data while developing isolated, single-field tools. This paradigm faces three critical constraints: fragmented scientific data systems, insufficient specialized reasoning capabilities, and closed research ecosystems.

To overcome these challenges, twelve CAS institutes, including the Institute of Automation, the Computer Network Information Center, the National Science Library, the Academy of Mathematics and Systems Science, and the Institute of High Energy Physics, among others, jointly designed and developed ScienceOne.

The model has achieved a comprehensive understanding of complex scientific data modalities, including waveforms, spectra, and fields, while integrating core capabilities for literature extraction, knowledge reasoning, and orchestrating computational tools. Researchers can easily access the model capabilities at various stages of their scientific work, enabling seamless integration of AI into scientific research.

Built upon China's open-source foundation models with deep scientific customization, ScienceOne integrates a suite of proprietary models developed for common scientific data types, while incorporating domain-specialized tools such as AlphaFold and MatterGen.

It demonstrates systematic mastery of core principles, laws, and specialized knowledge across fundamental disciplines including mathematics, physics, chemistry, astronomy, earth science and biology.

Evaluations have confirmed its state-of-the-art performance in mathematics, physics, chemistry, materials science and biology, with leading capabilities in scientific tool invocation and reasoning. ScienceOne further attained excellent results in Humanity's Last Exam (HLE).

Based on ScienceOne, the research team also developed two scientific intelligent agents.

One is designed to assist researchers in thoroughly reading articles, writing reviews, and evaluating research topics and technical pathways. It has access to 170 million pieces of scientific literature and real-time open-source scientific information. With its support, literature review tasks that previously required 3 to 5 days can now be completed in just 20 minutes.

The other agent aims to lower the barrier to using scientific tools, allowing for autonomous planning of over 300 scientific computing tools. It can automatically identify research tasks and intelligently schedule the optimal tools, thereby enhancing the efficiency of the research process.

Accelerating discovery across disciplines, ScienceOne currently powers the X-Cell digital cell platform automating biological target identification, enhances particle simulation efficiency at the Beijing Electron-Positron Collider, improves molecular prediction accuracy in chemistry, streamlines global telescope coordination for astronomy, and supports structural design innovation for high-speed rail systems.