Introduction
The Beijing Academy of Artificial intelligence, more formally known as the Beijing Academy of Artificial Intelligence (BAAI) has emerged since its founding in 2018 as a prominent and ambitious research institution in the global artificial intelligence landscape. Supported by both municipal and central government authorities, BAAI pursues fundamental research in machine learning, large-scale models, multimodal systems and embodied intelligence. Its work spans theoretical foundations, practical architectures and open-source ecosystems designed to accelerate both scientific inquiry and industrial adoption. This paper explores the institutional background of BAAI, its principal research thrusts, including large language and multimodal models, embodied intelligence and open infrastructure and the broader scientific, ethical and geopolitical contexts within which the academy operates. By analysing research outcomes, strategic priorities, collaborative structures and open science contributions, we assess BAAI’s role in advancing artificial intelligence and shaping research paradigms in the twenty-first century.
Artificial intelligence research; broadly defined as the scientific study of computational systems capable of perception, reasoning, learning and action stands at a pivotal moment due to rapid advances in deep learning, neural architecture search, life-long learning and embodied artificial intelligence systems. Within this dynamic scientific environment, the Beijing Academy of Artificial Intelligence (BAAI) has gained notoriety for its strategic focus on foundational research and original innovation rather than narrow applied engineering. It occupies a unique position in China’s artificial intelligence ecosystem, linking government policy, academic research and industrial innovation in an integrated, open-science framework.
This paper aims to provide an in-depth examination of BAAI’s research on artificial intelligence, emphasising core scientific contributions, organisational strategy and implications for global AI research. We first characterise BAAI’s institutional context and research objectives, then review its major research outputs and scientific programmes. We conclude by situating BAAI in a broader context of global artificial intelligence dynamics, ethical challenges and research governance.
Institutional Background and Mission
Established in November 2018, BAAI was founded under the auspices of Beijing’s municipal science and technology authorities, with the aim of creating a new type of non-profit research organisation prioritising original exploration in artificial intelligence. According to official statements, the academy was designed to address “major core foundational theoretical problems” in artificial intelligence by assembling top researchers from universities, research institutes and industry. The strategic intent was to establish Beijing as a global hub for artificial intelligence innovation and to support original scientific breakthroughs; a departure from more commercially oriented approaches that dominated earlier phases of China’s artificial intelligence development.
BAAI is governed by a council of academic and industry leaders and operates with support from both the Beijing Municipal Government and the Ministry of Science and Technology. It promotes a flexible research structure that fosters interdisciplinary collaboration and rapid iteration across projects. The institute explicitly positions itself as a hub for both theoretical and applied work, with an emphasis on open research and the cultivation of scientific talent.
BAAI’s core mission encompasses several interrelated goals:
1. Advance fundamental theory in artificial intelligence, including foundational principles for learning, reasoning and representation.
2. Develop large-scale models and systems that push the boundaries of current AI capabilities, particularly through scalable architectures and multimodal learning.
3. Promote open science and community engagement, including open platforms for model training, datasets and evaluation tools.
4. Bridge academia, industry and public policy by fostering collaboration and addressing ethical, safety and governance issues in artificial intelligence research.
These objectives reflect an integrated approach that leverages both intellectual exploration and ecosystem building to achieve scientific impact and societal relevance.
Research Agenda and Scientific Domains
BAAI’s research agenda spans a range of topics within artificial intelligence, including large language models (LLMs), multimodal systems, embodied intelligence, open-source artificial intelligence infrastructure and artificial intelligence safety. Below, we examine each of these domains in detail.
Wu Dao and Large Multimodal Models
One of BAAI’s earliest and most internationally visible research outputs is the Wu Dao series of models. Wu Dao is a multimodal pre-trained intelligence system designed to handle and integrate diverse data modalities; notably text and images, within a unified learned representation.
Wu Dao 1.0 debuted in early 2021, marking China’s first large-scale intelligent model and drawing wide academic and media attention for its ambitious scale and multimodal design. Its immediate successor, Wu Dao 2.0, announced later in 2021, contained approximately 1.75 trillion parameters, an order of magnitude larger than widely recognised contemporaries such as GPT-3 (175 billion parameters) and was trained on a diverse corpus comprising terabytes of images and text.
The multimodal architecture of Wu Dao is intended to enable integrated language and vision tasks, such as generating text from images or producing photorealistic images from text prompts. It illustrates an explicit research strategy aimed not merely at scaling up transformer-style models but also at exploring architectures capable of integrating and reasoning across data modalities.
In addition to natural language generation and image synthesis, Wu Dao variants have been applied to tasks such as protein structure prediction, an area historically dominated by bio-oriented artificial intelligence like DeepMind’s AlphaFold, demonstrating the flexibility and broad applicability of large multimodal models.
Open Infrastructure and Research Platforms
Beyond single models, BAAI has contributed to artificial intelligence research infrastructure by developing open-source tools and platforms designed to support the training, deployment and evaluation of large systems.
FlagAI is an extensible toolkit for training and deploying large-scale models across various tasks and modalities. Its design supports both research experimentation and practical deployment, enabling broader participation in AI research. Importantly, FlagAI has been accepted as an incubation project within the Linux Foundation, underscoring its potential as a community-driven foundation for artificial intelligence infrastructure.
The Jiuding computing platform supports artificial intelligence researchers by providing massive computational capacity (reportedly on the order of petascale equivalent) and cross-architecture compatibility for diverse AI workloads. Designed to facilitate experimentation with large models, Jiuding reflects BAAI’s emphasis on building shared infrastructure foundations that reduce barriers to high-performance artificial intelligence research.
The emphasis on open and scalable infrastructure contrasts with proprietary and commercially proprietary systems, emphasising community participation and reproducible research, a hallmark of BAAI’s intended contribution to global artificial intelligence science.
Embodied Intelligence and Robotic Cognition
Artificial intelligence research is increasingly attentive to embodied artificial intelligence, systems that perceive, understand and act within physical environments, an area that connects perception, planning and control. BAAI’s embodied intelligence initiatives reflect this trend, with substantial work on both simulation and real-world robotic cognition.
Recent published work from BAAI researchers has introduced RoboBrain models that unify vision, language and planning capabilities within an embodied context. According to technical reports, RoboBrain 2.0 presents a vision-language foundation model with heterogeneous architectures that integrate visual encoders and language modules, achieving competitive performance on diverse embodied reasoning tasks such as spatial understanding, long-horizon decision-making and dynamic interaction in physical environments.
Related experimental work describes the integration of multi-modal data and robotic manipulation capabilities, such as trajectory prediction and affordance perception, in foundation models designed for robotic control and interaction. This body of research illustrates a methodological shift from disembodied language tasks toward integrated physical cognition and situational reasoning.
Complementary to foundation models, BAAI has developed operating systems and modular frameworks, for example, RoboOS 2.0, that support deployment of embodied intelligence across robotic platforms. These software systems provide modularised infrastructures capable of integrating sensory perception, task planning and motor execution on diverse hardware, reducing the real-world deployment gap for embodied agents.
In total, embodied intelligence research at BAAI demonstrates a concerted effort to bridge abstract representations and concrete actions, a frontier that many consider essential to the development of more general forms of artificial intelligence.
Scientific Applications in Biology and Medicine
An emergent branch of BAAI’s research portfolio focuses on applications of artificial intelligence to scientific domains including biology and medicine. Beyond language and robotics, models trained on scientific corpora are being applied to protein structure prediction, drug design and biological simulation.
For instance, OpenComplex2, a model targeting complex biological molecules, extends static structural prediction toward dynamic conformational modelling. This capability promises to shorten research cycles in drug development and toxicity evaluation. Partnerships with biomedical research institutions and hospitals indicate a growing emphasis on cross-disciplinary work that harnesses AI to accelerate scientific discovery.
Open Science, Collaboration and Community Building
BAAI’s research strategy is notable for its emphasis on open science. By releasing models, toolkits and datasets under permissive licences, the academy has fostered a global community of researchers and developers who can experiment with state-of-the-art models without prohibitive costs. For example, BAAI reports that its open-source LLMs have been downloaded tens of millions of times worldwide, suggesting broad engagement beyond China’s borders.
Annual conferences organised by BAAI bring together artificial intelligence scholars, practitioners and policymakers from around the world, facilitating knowledge exchange on topics ranging from multimodal learning to autonomy and safety. These events reinforce scholarly discourse and contribute to a pluralistic artificial intelligence research culture.
Collaborations also extend to academic institutions, research laboratories and industry partners, both domestic and international, reflecting BAAI’s role as a nexus between scientific investigation and technological application.
Ethics, Governance and Geopolitical Context
Artificial intelligence research, particularly at scales pursued by institutions like BAAI carries significant ethical and governance implications. Questions about safety, robustness, dual-use concerns and equitable access to technology are central to contemporary discourse on artificial intelligence research governance.
BAAI’s research ecosystem includes engagement with standards, ethical frameworks and policy dialogues aimed at responsible development and deployment of artificial intelligence technologies. However, its alignment with national strategic objectives and geopolitical priorities renders its activities part of broader international debates about AI leadership, technological sovereignty and regulatory harmonisation.
In 2025, for example, BAAI was added to the U.S. Department of Commerce’s Entity List, highlighting the ways in which geopolitical competition intersects with scientific research in sensitive technologies. While such developments reflect international tensions rather than scientific merit per se, they emphasise the geopolitical significance attributed to institutions that drive frontier research.
Strategic Significance and Future Direction
BAAI’s research portfolio exemplifies a multi-pronged approach to artificial intelligence that emphasises both foundational model innovation and systemic infrastructure development. Its large language and multimodal systems contribute to expanding the frontiers of scalable artificial intelligence architectures, while embodied intelligence initiatives advance integrated perception-action systems that bridge simulation and real-world interaction.
BAAI’s work is significant not only for its technical achievements but also for its strategic orientation toward open science and ecosystem building. By lowering barriers to high-end computation and model access, the academy amplifies research productivity and fosters collaboration across institutions. Moreover, its participation in ethical and governance dialogues signifies a recognition of broader societal impacts attendant to artificial intelligence.
Looking forward, continued research at BAAI will likely push further into areas such as neural computation for scientific discovery, generalisation across modalities and the development of more robust and adaptive embodied agents. Its position as a hub for talent and infrastructure suggests that it will remain influential in shaping both China’s AI research landscape and broader global scientific agendas.
Conclusion
The Beijing Academy of Artificial intelligence, understood here as the Beijing Academy of Artificial Intelligence (BAAI), represents a major research endeavour in contemporary artificial intelligence. Its work spans large-scale models, multimodal architectures, embodied intelligence frameworks and cross-domain scientific applications. Through open-source releases, collaborative networks and strategic research infrastructure, BAAI contributes to the global scientific community while navigating the complex ethical and geopolitical context of twenty-first-century artificial intelligence.
In sum, BAAI’s research trajectory reflects an ambitious pursuit of both scientific depth and societal relevance, situating the academy as a central player in the ongoing evolution of intelligent machines.