China Releases Upgraded Chemical AI Model to Automate Industrial Tasks

Chinese researchers and tech firms have unveiled Intelligent Chemical Large Model 3.0Pro, an upgraded vertical artificial intelligence system tailored for the chemical industry. Developed by the Dalian Institute of Chemical Physics under the Chinese Academy of Sciences alongside iFlytek and Alibaba Cloud, the model expands beyond question answering to plan and execute complex industrial workflows. Incorporating more than 400 chemical tools and agents, the system achieved over 80 percent accuracy in benchmark evaluations and is backed by a full-chain data center and an industrial pilot facility.

Chemical engineering involves intricate physical constraints, hazardous operating conditions, and complex multi-step processes that general-purpose foundation models struggle to handle reliably. By integrating specialized autonomous agents with real-world operating data, vertical systems like 3.0Pro aim to compress research and development timelines for novel materials and catalysts while mitigating scaling bottlenecks between laboratory experiments and factory production lines.

A consortium led by the Dalian Institute of Chemical Physics under the Chinese Academy of Sciences, alongside Chinese technology firms iFlytek and Alibaba Cloud, has launched the Intelligent Chemical Large Model 3.0Pro, state media reported on August 31. The system represents the latest iteration of China's first domain-specific foundation model for the chemical sector, according to Xinhua News Agency. The upgrade marks a transition from standard technical question-and-answering capabilities toward autonomous planning and execution of specialized chemical engineering tasks. General foundation models frequently struggle in the chemical sector due to rigorous process constraints, safety margins, and complex multi-variable physics. To address these hurdles, the developers constructed a four-tier architecture combining cognitive natural language capabilities with operational agent execution. The platform integrates more than 400 chemical agents and digital tools trained on specialized domain datasets, establishing a closed-loop workflow spanning comprehension, planning, execution, and verification. In official evaluations, the 3.0Pro model recorded an accuracy rate of 81.96 percent on technical text question-and-answering benchmarks and 80.75 percent on multimodal assessments, reflecting what developers described as a marked performance gain over prior builds. Before this release, earlier iterations had already been made available to more than 300 commercial enterprises, universities, and research institutions, accumulating over 14 million calls via its application programming interface (API). To overcome persistent shortages of high-fidelity operational data from actual plants, the engineering consortium established a petrochemical full-chain big data center paired with a thousand-ton-scale intelligent pilot testing platform. The physical facility provides empirical validation and operational telemetry to refine the model's predictive and optimization capabilities. Looking ahead, the development team intends to advance the project toward version 4.0. Project planners indicated that subsequent releases will focus on bridging remaining operational gaps between laboratory-stage synthesis and full-scale factory production.