Driven by a vision to cultivate globally competitive AI leaders and accelerate original technological breakthroughs, the Zhongguancun Academy (ZGCA) and the Zhongguancun Institute of Artificial Intelligence (ZGCI) have just released eight key areas of research progress since our founding in September 2024. These breakthroughs span cutting-edge academic domains, including AI Core (AIcore), AI for Natural Sciences (AI4S), and AI for Social Sciences (AI4SS) — reflecting our commitment to interdisciplinary innovation and real-world impact.
Super Software Intelligence (Modelware): Defining New Boundaries of AI Capabilities in the Software Field
The Super Software Intelligence (Modelware) technology developed by Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence, based on a novel model-architecture design and centered on analysis of program logic, data structures, and execution behavior, has for the first time achieved deep penetration into the underlying operating mechanism of software，greatly extending the capability boundaries of AI in the software field. Using Modelware technology, the team has implemented pioneering features such as automatically and seamlessly porting binary programs from one platform to another without manual coding; automatically transforming traditional software into MCP services or AI agents; and automatically modifying or generating software on demand based on natural-language instructions.
MegaDFT: Scaling up Kohn-Sham DFT to 100,000 atoms
Density Functional Theory (DFT) is currently the most widely used quantum chemistry method. Kohn-Sham DFT (KS-DFT) plays a critical role in computational analysis across chemistry, life sciences, and materials science. However, the computational complexity of KS-DFT scales cubically with the number of atomic orbitals, making it prohibitively expensive. Existing software typically handles systems with only up to a few thousand atoms.MegaDFT, developed by ZGCA·ZGCI, has for the first time achieved hybrid-functional-level KS-DFT calculations for molecular systems containing up to 100,000 atoms, scaling the target molecular system size by two orders of magnitude. With MegaDFT, it is now possible to perform hybrid-functional DFT calcuations for large biomolecular and non-periodic material systems. For example, it enables quantum chemical analysis with high accuracy for over 99% of protein structures in PDB, unlocking quantum insights for ultra-large biomolecular systems.
RLinf: The World’s First Open-source Large-Scale Reinforcement Learning Framework for Embodied Intelligence
Reinforcement learning is a key technology driving the next generation of embodied intelligence. However, large-scale reinforcement learning frameworks tailored for embodied intelligence are still absent worldwide, which has severely hindered rapid iteration in this field.The ZGCA&ZGCI– Infinigence AI Infra Lab in collaboration with Tsinghua University, has launched RLinf, the first open-source reinforcement learning framework for embodied intelligence that integrates simulation, training, and rollout.RLinf features three major innovations:
1. Hybrid fine-grained pipeline execution mode designed for the unique requirements of simulation–training–rollout integration in embodied intelligence, boosting system efficiency by 120% and improving model performance by 40%–60%.
2. Unified worker-based programming interface tailored to the distinct training needs of the cognitive brain (high-level decision-making) and motor brain (low-level control). It leverages micro-execution flows to implement macro workflows, achieving a balance of usability, efficiency, and flexibility.
3. First-ever second-level GPU online auto-scaling mechanism for large-scale cluster training, significantly enhancing training stability.Together with multiple partners, we are building an open-source ecosystem for embodied intelligence centered on RLinf, paving the way for the sustainable and robust development of embodied intelligence in China.
The World’s First AI System with One Billion Agents for Earth-Scale Social Simulations
In social life, every piece of news and every policy profoundly affects millions of lives and ripples across countless industries. Zhongguancun Academy has developed the world’s first AI system with one billion agents, providing an earth-scale simulation platform for social science research, policy modeling, and decision forecasting.The system is built on algorithmic and systemic innovations working in tandem, overcoming the long-standing "trilemma" in social simulation, namely large-scale systems, high accuracy, and low computational cost. It introduces a pioneering heterogeneous-model intelligent routing engine that reduces computational costs by three orders of magnitude with less than 1% loss of accuracy, enabling tens of thousands of inferences per second. Grounded in massive real-world datasets, the system equips agents with unique identities, cognitive abilities, social relationships, and behavioral patterns, thereby reproducing complex socio-economic phenomena.It has already been applied to media analysis at China Media Group, where, by leveraging social media data, it accurately simulated public opinion involving tens of millions of individuals, providing more scientific guidance for content creation.
Discovery Virus (DiscoVir): High-Precision, Fine-Grained Virus Discovery to Unveil the 'Viral Dark Matter' and Build a Protective Shield for Humanity
Abstract: Despite the existence of tens of billions of virus species in nature, only around ten thousand have been studied and classified. To advance our understanding of the viral universe, the ZGCA·ZGCI have developed Discovery Virus (DiscoVir) — an accurate and fine-grained virus discovery model. DiscoVir enables systematic classification from domain to family level and can detect potential novel viral phyla.Leveraging DiscoVir, the team has identified over 1.34 million previously uncharacterized viral sequences from global metagenomic datasets. These sequences—referred to as “viral dark matter”— have been annotated and categorized, significantly expanding our knowledge of global viral diversity. This breakthrough provides a foundation for research in virology, evolutionary biology, and public health.
AI Drives Breakthrough in Gene Editing Technology, Building a Sovereign and Controllable Future
Gene editing technology holds significant value in fields such as bio-breeding and healthcare. However, China currently lacks a sovereign intellectual property system for high-efficiency gene-editing enzymes, presenting a substantial "chokehold" risk. The Zhongguancun Academy & the Zhongguancun AI Institute have engaged in close collaboration with the Chinese Academy of Agricultural Sciences (CAAS) to jointly tackle this challenge.The team has adopted a dual-track innovation path. On one hand, it utilizes AI-driven directed evolution technology, starting with low-efficiency gene-editing enzymes discovered in China, to rapidly iterate and enhance their editing efficiency and practicality. On the other hand, based on an AI-powered deep understanding of the protein "sequence-to-structure" relationship, the team starts with given high-efficiency gene-editing enzymes and designs entirely new proteins by progressively increasing sequence diversity while maintaining their structure and function, thereby securing independent intellectual property.The team has successfully developed over 90 novel gene-editing enzymes that are compact and highly stable. Wet lab experiments have shown that the highest editing efficiency has surpassed the critical 60% threshold and is continuously being optimized. This work lays a solid technological foundation to progressively build a sovereign and controllable intellectual property system for gene-editing enzymes.
Uncharted Territory of AIDD: Tackling Diseases with hard-to-drug target and unknown target
Mainstream AI-driven drug discovery technologies are predominantly designed for diseases with well-defined molecular targets. As a result, approximately 95% of diseases—characterized by either “undruggable targets” or the absence of clear targets—remain largely untouched, forming a vast “no-man’s land” in AI pharmaceutical research.To systematically address this challenge, ZGCA·ZGCI have developed a cross-domain biological foundation model capable of comprehensively understanding disease mechanisms through multi-modality data integration, thereby reducing reliance on predefined targets.For diseases with undruggable targets, the team focused on malaria and successfully designed novel drug candidates capable of binding to disordered target proteins in Plasmodium species. These molecules demonstrated exceptional activity in wet-lab experiments. For diseases with unclear targets, the team focuses on fibrosis. Using cellular phenotypes as input, they designed a series of new drug candidates that significantly outperformed reference compounds in terms of biological activity.These results highlight the potential of cross-domain biological foundation models to unlock new pathways in AI-driven drug discovery, offering promising strategies to explore previously inaccessible therapeutic landscapes.
AI Understanding the Physical World
While current AI models have achieved near-human performance in text-based reasoning, they still struggle to understand the real physical world. Even the most advanced large language models show significant gaps compared to humans in complex multimodal physical reasoning tasks. A joint team from Zhongguancun Academy and Peking University has developed a groundbreaking multimodal reasoning framework. By enabling deep interaction between a "Physical Scene Captioner" and a "Knowledge Reasoner," the system achieves profound understanding and precise solutions to physical problems. This approach has significantly outperformed models such as Gemini-2.5 and ChatGPT-o3, claiming the championship in the ICML 2025 SeePhys Challenge. Looking ahead, the team plans to build a "comprehensive evaluation" system spanning multiple disciplines, incorporating adaptive dynamic difficulty adjustments to drive the continuous evolution of AI reasoning capabilities. This effort will usher in a new era, transitioning AI from the text-based world to comprehensive perception and understanding of the real world.
“AI Future” Hackathon Wraps Up After 12-Hour Challenge