Shenzhen’s 11B-parameter science model debuts in Shanghai, integrating six data types to bridge fields from DNA to weather

On July 19, at the Scientific Intelligence Open Forum during the 2026 World Artificial Intelligence Conference, the Shanghai Institute for Scientific Intelligence introduced a foundational super brain: the Shenzhen Scientific Multimodal Foundation Model. Designed to support multidisciplinary scientific research, the model draws its name from the "Divine Pearl Iron" in Journey to the West. The development team aims for this open model to handle diverse scientific tasks efficiently, enabling broader researcher participation in testing, usage, and collaborative development. It serves as the "Great Sage" system-level intelligent research agent, launched in early March alongside its super brain.

Technically, Shenzhen integrates disciplines including material science, life science, and earth science, achieving a stage-wise unified representation of cross-domain scientific knowledge and multimodal understanding. With approximately 11 billion parameters, the model supports scientific understanding and multi-type result generation for six distinct data categories: DNA, RNA, proteins, small molecules, Earth systems, and medical images, all within a single framework. Utilizing Qwen3-VL-8B as its shared backbone, the model employs dedicated data processing paths for each of the six scientific data types. Prior to connecting to the shared model, it preserves sequence order, molecular connection structures, meteorological field spatial distribution, and local details of medical images. Practically, Shenzhen can directly generate RNA sequences, computer-readable molecular representations (SMILES), global meteorological fields, and medical image segmentation results, effectively integrating understanding and generation capabilities into one cohesive framework.

Open source and accessibility are central to this model. The Shanghai Institute for Scientific Intelligence has released model weights, inference code, example scripts, and usage documentation. Researchers can download the model or access APIs via the Xinghe Qizhi Scientific Intelligence Open Platform, utilizing it alongside over 1,500 scientific models and tools available on the platform. Additionally, weights are available on Hugging Face, with code hosted on GitHub, fostering a collaborative open scientific intelligence community.

This release occurred during the Scientific Intelligence Open Forum at WAIC2026. Guided by the Office of the World Artificial Intelligence Conference Organizing Committee and hosted by Fudan University and the Shanghai Institute for Scientific Intelligence, the forum focused on scientific closed-loop discovery in the AI-native era. It brought together renowned scientists and young representatives from industry, academia, and research to discuss how artificial intelligence could restructure the entire scientific discovery process at a methodological level, aiming to create an open and collaborative scientific intelligence ecosystem. Four distinguished scientists delivered keynote speeches, offering diverse perspectives on scientific intelligence.

Professor Arie Y. K. Wachter, Nobel Prize in Chemistry laureate and professor at the Chinese University of Hong Kong, Shenzhen, emphasized that AI must be grounded in reliable physical mechanisms. Professor Gilles Brassard, Turing Award winner and professor at the University of Montreal, expressed high expectations for quantum computing's potential in areas like new material discovery, advising young researchers to pursue their interests rather than industry trends, as current research findings may only fully materialize after a decade.

Wang Jian, Director of the Zhejiang Lab and academician of the Chinese Academy of Engineering, shared his views on scientific foundation models and science as a whole. He noted that scientific intelligence resides in data rather than paper texts, yet current foundation models remain text-based. He argued that scientific data should become the native inhabitants of scientific intelligence, meaning data from various disciplines should be tokenized and integrated into a common representation space. He believes AI is becoming as fundamental as mathematics, shifting the research paradigm from STEM to STE+MAP—the integration of mathematics, AI, and public infrastructure—thereby unifying science.

Professor Jianqing Fan, member of the U.S. National Academy of Sciences and professor at Princeton University, discussed intelligent science and society, defining AI as a dynamic cycle of statistical learning and optimization decision-making. He explained how AI transforms data into social insights, citing examples such as socioeconomic measurement, financial risk control, and large model applications. He emphasized that while AI serves socioeconomic development and promotes intelligent agents and scientific innovation, it must continuously address challenges such as changes in employment structure, educational transformation, ethical security, and the maintenance of effective human control.

During the summit dialogue, participants reached a consensus: a key factor for original innovation in the AI-native era is whether people, data, models, experiments, and academic judgment can form a new collaborative discovery mechanism. This requires not only systematic upgrades in research infrastructure but also a fundamental transformation in the thinking patterns and capability structures of a new generation of researchers.

Qi Yuan, specially appointed professor at Fudan University and director of the Shanghai Institute for Scientific Intelligence, outlined the ultimate vision of this transformation. He stated that AI should evolve from predicting the next token to discovering unknown laws, with the discovery of these laws marking the beginning of super intelligence. The core of this process involves compressing high-dimensional space to uncover concise principles and achieving an efficient scientific verification loop, promoting the integration of digital and physical worlds and the emergence of new model architectures. He reminded attendees that real research involves a long chain process including literature, hypotheses, data, models, simulations, experiments, and feedback, and that the field of scientific intelligence is currently focused on developing systemic capabilities to organize complex research processes.

As Shenzhen, with its 11 billion parameters, consolidates six types of scientific data into a single mind, the form of scientific discovery may be quietly rewritten.