China's Open-Source AI System Identifies Particle Decay Modes, Reshapes Scientific Research
An AI system released by China's Academy of Sciences in April has identified more than 11 previously unknown particle decay modes-discoveries that would normally require research teams to spend years sorting through vast amounts of high-background noise data. In space science, another AI agent in the same system has achieved 100 percent accuracy in predicting X-class solar flares while operating an autonomous telescope observation system.
The system, called ScienceOne 100, is not a chatbot or image generator. It is purpose-built for scientific research across mathematics, physics, biology and other disciplines. Unlike general-purpose large language models, it is designed to understand scientific theories, conduct research and enable practical implementation.
What distinguishes ScienceOne 100 is that the Academy of Sciences has made it openly available to researchers worldwide rather than keeping it proprietary.
General models fail at rigorous science
Popular large language models work well for everyday tasks but fall short in scientific research, according to Zhang Jiajun, a researcher at the Academy's Institute of Automation. They are prone to factual errors and reasoning hallucinations, and they cannot interpret scientific data such as particle signals, astronomical spectra or cellular information.
ScienceOne 100 is designed to progress through three stages: from a research tool to a research collaborator, and ultimately to a virtual AI scientist capable of independently completing full research cycles-data collection, analysis, hypothesis formulation, verification and iteration.
The system emerged from joint research coordinated by dozens of Academy-affiliated institutes. It advances AI for science from fragmented, closed operations to collaborative, open platform-based development, fundamentally changing how traditional scientific research is organized and conducted.
China's open-source strategy diverges from Western approach
While many Western technology companies keep their most advanced models proprietary, Chinese companies have embraced open source. Nearly all of China's top-performing AI systems are widely available.
DeepSeek released a preview version of its next-generation V4 model planned to be fully open-sourced. Multiple Chinese technology companies have upgraded their open-source ecosystems in recent months. According to Hugging Face, a major open-source community, Chinese models accounted for 41 percent of downloads in the past year, surpassing the United States.
This approach reflects what experts call "Chinese-style open source"-one that combines code sharing with the principle that "shared prosperity is true prosperity" alongside China's strategy of achieving technological self-reliance through cooperation rather than isolation.
Open-source models deliver five competitive advantages
Chinese open-source models possess five distinct strengths, according to Zhu Yue, an expert from Tongji University Law School:
- Performance: Models such as DeepSeek-r2 rank among the world's top performers, while most advanced American models remain closed-source.
- Openness: Beyond releasing model weights, core technical ideas and methodologies are shared through peer-reviewed papers, technical reports and other platforms.
- Ecosystem: Supporting frameworks, datasets and skill modules are made accessible alongside foundational models.
- Minimal restrictions: Most use open licenses like MIT with few additional constraints and loose application limits.
- Downstream applications: The models and supporting ecosystems enable numerous AI applications globally.
Regulation accompanies openness
Openness does not mean the absence of regulation. Security and copyright issues concerning open-source models remain significant challenges.
China has not released an official AI law draft, but law experts have drafted two influential proposals-Model AI Law and AI Law-that serve as references for understanding China's approach to AI regulation. Chinese experts argue that openness and regulation can coexist.
In a Science article published in October 2025, Chinese experts wrote that "in-depth global cooperation, for which openness has always been and will always be necessary," is essential. They called for solutions that simultaneously promote openness and innovation as "the most workable lowest common denominator for the global AI governance dialogue."
Real-world deployment across eight disciplines
ScienceOne 100 has already deployed large models across eight major scientific disciplines on research frontlines. More than 100 scenarios have achieved large-scale application across fundamental research, engineering, public welfare and national strategic priorities.
The Academy of Sciences has established in-depth cooperation with Belt and Road partner countries, providing access to ScienceOne 100 to empower their scientific research. The Academy has inked 120 government-to-government science and technology cooperation agreements, a significant portion with developing countries.
According to Zhang Jiajun, one scientist equipped with ScienceOne 100 will eventually accomplish research work that currently requires dozens or even hundreds of researchers. Through open-sharing mechanisms, these capabilities are increasingly becoming resources that benefit countries around the world.
For researchers looking to understand how AI systems are reshaping scientific discovery, explore AI for Science & Research learning resources.
Your membership also unlocks:
