Investors seeking to identify promising themes should focus on in-depth analyst research for authoritative, professional, timely, and comprehensive insights.
Beijing's status as a premier hub for artificial intelligence is being reinforced by a series of significant innovations, from hosting the first listed global large language model entity to launching the world's largest open-source model by parameter count.
On July 16, Beijing-based company Moonshot AI unveiled KimiK3, the world's first open-source model at the three-trillion parameter scale, engineered for advanced scenarios such as long-context programming, knowledge work, and reasoning. Beijing's capacity to support one of the world's largest models by parameter count is solid proof that its capabilities, from model training to final output, rank among the global elite.
Introduction to the 3-Trillion Parameter Open-Source Model
KimiK3 is a 2.8-trillion parameter model. Its development focus was not merely on scaling up parameters but on synergistic optimization of model architecture and training methods to translate scale advantages into enhanced capabilities. A representative from Moonshot AI explained that parameters are akin to neural connections in the human brain; nearly three trillion parameters mean the model can store more knowledge and patterns in its "mind," enabling it to know more, think deeper, and answer more accurately.
In evaluations released by Moonshot AI, KimiK3 demonstrated world-leading comprehensive abilities. It can not only execute specific code generation and modification tasks but also engage in open-ended creation and complex knowledge production. In digital content generation, the model can combine 3D reasoning, programming, and visual capabilities to transform concepts, images, or videos into interactive experiences. In scenarios like financial research and industrial analysis, it can organize and analyze corporate documents, industry data, and long-term sector materials to produce readable, verifiable, and checkable in-depth reports.
The Moonshot AI representative disclosed that the company is currently collaborating closely with inference partners and open-source maintainers to align technical details, ensuring the model can be reliably deployed across the ecosystem. The complete model weights will be released before July 27, 2026. Further details on architecture, training, and evaluation will be published alongside the KimiK3 technical report.
The Path to a Leading Global Position
"When a model's parameters reach the trillion level, it essentially enters the global first tier," stated Shao Bin, Director of the Zhongguancun Artificial Intelligence Research Institute, emphasizing the significance of Moonshot AI's latest ultra-large parameter model. For a considerable period, the "size" of large models primarily reflected scale expansion, following the scaling laws—where continuous increases in computing power, model size, and data volume lead to corresponding leaps in intelligence levels. With ongoing investment in computing power and intensified competition in specific fields, future models may see their parameter counts expand even further.
Beijing's ability to support one of the world's largest models by parameter size serves as full confirmation that its processes, from model training to final output, are positioned within the global top tier.
Global Open-Sourcing of a Meteorological Model
Beijing has consistently been at the forefront of AI development: it has produced a series of internationally leading models like GLM-5.2 and KimiK3; it has registered and launched 259 large models, ranking first nationally; and it is home to the "first listed global large language model entity." This year, Zhipu AI successfully listed in Hong Kong, achieving that status.
Wang Shaolan, President of the company, explained that "enabling machines to think like humans" has been the direction Zhipu AI chose from its inception. In 2025, the company released GLM-4.5, GLM-4.6, GLM-4.7, and GLM-5, with all four generations of its foundational models ranking among the top global open-source offerings, gradually acquiring reasoning abilities approaching human levels. In June 2026, GLM-5.2 was released, with its coding capabilities ranking first among global open-source models. Concurrently, user adoption of GLM-5.2 continues to grow worldwide, marking a phase of rapid development in model application.
On July 17, at the 2026 World Artificial Intelligence Conference (WAIC) Meteorology Session, the China Meteorological Administration unveiled the "Fenghe" large language model for AI meteorological services and launched the "Fenghe" Global Open Source Initiative. Fenghe is the world's first open-source meteorological large language model at the hundred-billion parameter level, jointly developed by the China Meteorological Administration Public Meteorological Service Center and Zhipu AI, among others. It provides intelligent support for weather analysis, risk assessment, and meteorological services.
The launch of the Fenghe Global Open Source Initiative represents more than just open-sourcing model code; it is a complete technical delivery solution allowing global users to integrate Fenghe into various terminals like embodied AI, apps, and mini-programs, thereby building an open and shared meteorological AI ecosystem to make weather AI services accessible to everyone.
"This case in the meteorological public service sector demonstrates that Zhipu's technology has proven effective in real-world scenarios, which in itself is positive confirmation of its application value," Wang Shaolan noted. Currently, the GLM series of large models is applied in over 20 industries, including public governance, industrial manufacturing, energy and power, finance, and the internet, accelerating the march toward the AGI era.
Wang Shaolan believes that global research teams and researchers should maintain an open and collaborative attitude rather than keeping technology closed. While continued effort is needed for technological breakthroughs, areas like safety governance, model interpretability, and inherently embedding safety into models equally require international cooperation.
Replicable AI Application Experience Emerges in Beijing
In the field of artificial intelligence, Beijing has established a tripartite development framework characterized by continuous breakthroughs in technological innovation, a highly concentrated talent pool, and deepening industrial implementation.
Shao Bin explained that technological breakthroughs stem from both academic forces like universities and research institutes, and corporate entities such as Zhipu AI and Moonshot AI, with all parties continuously driving technological progress. Simultaneously, talent is a key factor in AI competition. Beijing boasts the nation's densest concentration of research institutes and national strategic scientific research forces, giving it a significant talent advantage. Furthermore, the AI industry chain is highly完备, providing excellent ecosystem support for AI to empower countless industries.
It is worth noting that the application forms of the large model industry are evolving from past conversational modes towards action-oriented models that genuinely介入 production processes. Reports indicate that Beijing has formed a batch of replicable and promotable AI application experiences in key areas such as healthcare, education, government affairs, and industry.
Among these, Beijing has pioneered the "AI + New Materials" innovation paradigm, with a series of integrated innovation成果 accelerating their emergence, providing reference and借鉴 for related industry development.
Shao Bin, Director of the Zhongguancun Artificial Intelligence Research Institute, stated that a core reason Beijing can successfully advance specialized initiatives like AI+ New Materials lies in deeply binding technology demanders with technology providers. By injecting real-scenario data, models evolve from possessing通用能力 to being able to complete specific professional tasks. Technology does not remain confined to the laboratory; instead, developers directly进驻 user scenarios to understand on-site what problems exist and what technologies are needed to solve them. This way, models gain both通用能力 and专业能力, transforming from "capable" models into tools that "know how to work."
A responsible official from the Beijing Municipal Development and Reform Commission stated that Beijing will下一步持续完善 its policy support system, optimize the innovation ecosystem, accelerate the cultivation of new forms of the intelligent economy, and expedite the construction of a global AI innovation highland with technological策源力 and industrial competitiveness. Efforts will focus on application scenario development,深入推动 the construction of application bases and标杆 scenarios, accelerating the formation of application demonstrations in key fields and industries, significantly提升 the performance and普及率 of智能 agents and智能 terminals, and promoting the coordinated development of the芯片, model算法, cloud, edge, and application quintet.
Comments