The ascension of Zhipu AI to the status of the world’s first publicly traded large-scale model corporation on January 8, 2026, represents a seminal juncture in the history of computational linguistics and artificial general intelligence (AGI). Established in 2019 as a technological manifestation of decades of research within Tsinghua University’s Knowledge Engineering Group (KEG), Zhipu AI has navigated a complex trajectory from academic curiosity to commercial dominance. This report delineates the organization’s evolution through its foundational academic rigor, the iterative development of its General Language Model (GLM) architecture, its pioneering work in multimodal synthesis, and its eventual conquest of the global capital markets via the Hong Kong Stock Exchange. The Academic Foundation: Tsinghua KEG and the Pedigree of InnovationThe narrative of Zhipu AI is fundamentally a narrative of academic continuity. The company did not emerge from the traditional venture capital incubator model but was instead "born in the lab," specifically the Knowledge Engineering Group (KEG) of the Department of Computer Science and Technology at Tsinghua University. Founded in 1996, KEG stands as one of the most venerable institutions in China for the study of natural language processing, knowledge graphs, and data mining. The leadership of Zhipu AI is comprised of some of the most cited scholars in the field of computer science. Chief Scientist Tang Jie, a professor at Tsinghua and Vice President of the Beijing Academy of Artificial Intelligence (BAAI), exemplifies the high-academic standard of the organization. With an H-index of 154 and over 110,000 citations across 1,300 papers, Tang Jie’s research in social network analysis and cognitive graphs provided the theoretical bedrock for Zhipu’s “data and knowledge dual-driven” approach. The presence of Li Juanzi and Xu Bin, both distinguished professors at Tsinghua, as “acting in concert” parties underscores the deep integration between the university’s research goals and the company’s commercial strategy. This relationship ensured that Zhipu AI remained at the bleeding edge of theoretical developments while having the structural capacity to operationalize them. From AMiner to the Dual-Driven Cognitive Engine The technological precursor to Zhipu AI's large models was AMiner, a global academic social network and intelligence system developed under Zhang Peng's guidance.[4, 6] AMiner served as an essential sandbox for handling massive, heterogeneous datasets—specifically scientific literature and academic profiles. This experience highlighted the limitations of pure statistical models, leading the founders to advocate for a "dual-driven" cognitive engine that fuses the massive data processing capabilities of deep learning with the structured reasoning of knowledge graphs.[6] This philosophical commitment remains a distinguishing feature of the GLM architecture, allowing it to navigate complex semantic tasks where data-only models often hallucinate. The Evolution of the General Language Model (GLM) Paradigm The core of Zhipu AI’s technical achievement lies in the GLM architecture. Unlike the Western-dominated transformer paradigms, which were initially bifurcated into auto-regressive models (like GPT) and auto-encoding models (like BERT), the GLM series sought to unify these objectives. The GLM-130B Breakthrough and Bilingual Sovereignty In 2022, Zhipu AI released GLM-130B, a bilingual (Chinese and English) model with 130 billion parameters. At the time, this was a significant achievement for the Chinese AI ecosystem, as it demonstrated that a domestic entity could train a model comparable to OpenAI’s GPT-3 on a massive scale while addressing the unique linguistic nuances of the Chinese language. The pre-training of GLM-130B utilized a multi-task objective that included span prediction and long-form generation. This flexibility allowed the model to excel in zero-shot and few-shot learning environments. Furthermore, the project was a collaborative effort involving the Ascend 910 AI processors, proving that the organization could optimize high-performance models for domestic hardware—a critical step for technical sovereignty. Democratization through the ChatGLM Series The launch of the ChatGLM series, particularly ChatGLM-6B, marked a moment pivotal in the "democratization" of large language models. By optimizing the model to run on consumer-grade GPUs, Zhipu AI empowered thousands of developers and small businesses to experiment with local LLM deployments. The iterative progress of the series is summarized in the following table: The ARC Framework and GLM-4.5: A Unified Foundation By August 2025, Zhipu AI introduced GLM-4.5, which redefined the metrics of a "Generalist Model" (Generalist Model) through the ARC framework: Agentic, Reasoning, and Coding.[15] The model’s design philosophy moved toward a "depth over width" structure, emphasizing more layers over hidden massive dimensions, a decision backed by research showing deeper models possess superior reasoning stability. Architecturally, GLM-4.5 utilizes a Mixture-of-Experts (MoE) configuration with 355 billion total parameters, of which only 32 billion are activated during any single inference step. This provides a favorable balance between the model's intelligence and its operational cost. Technical innovations such as Partial RoPE (Rotary Positional Embeddings), QK-Norm for numerical stability, and the Muon optimizer allowed the team to train on a massive corpus of 23 trillion tokens. Academic Mastery: Multimodal Synthesis and Specialized Intelligence Zhipu AI’s academic output is characterized by its breadth across multiple AI domains, particularly in the synthesis of images, video, and code. Image and Video Generation: The Cog Family The organization’s contribution to generative vision is spearheaded by the CogView and CogVideo series. CogView, presented at NeurIPS 2021, was a 4-billion-parameter transformer that utilized a VQ-VAE tokenizer to achieve state-of-the-art results in text-to-image synthesis. This work was foundational in proving that the transformer architecture, initially designed for text, could be effectively scaled for high-resolution visual generation. In the realm of video, CogVideo (presented at ICLR 2023) and the subsequent CogVideoX provided the first high-quality, open-source alternatives to proprietary video generators. CogVideoX, in particular, introduced large-scale movement support and improved temporal consistency, enabling it to generate videos with clearer, more stable frames that maintained identity across time. Coding Intelligence and the CodeGeeX Series The CodeGeeX project represents Zhipu's commitment to the "C" in ARC—Coding. CodeGeeX is a multilingual code generation model trained on more than 20 programming languages. Its architecture was specifically designed for integration into developer workflows, offering features such as code explanation, translation, and autonomous refactoring. By late 2024, CodeGeeX4-9B achieved the highest scores on the BigCodeBench for models under 20 billion parameters, demonstrating that the organization’s "compact but powerful" philosophy could outperform significan tly larger models in specialized domains. The Rise of Agentic Intelligence: AutoGLM and the "Operator" Concept In 2025, Zhipu AI transitioned its focus toward the development of AI Agents—autonomous systems that move beyond linguistic responses to take action within digital and physical environments. AutoGLM and the Technical Path to "Rumination" The release of AutoGLM marked the introduction of the first domestic "Deep Research" function and the transition of the LLM into an "Operator". The technical evolution of this system is distinct: it stems from the GLM-4 base model, evolving through the GLM-Z1 reasoning model into the GLM-Z1-Rumination (Contemplation) model. The "Rumination" capability allows the agent to simulate human reasoning through a continuous loop of exploration, verification, and correction. Unlike standard agents that operate via static APIs, AutoGLM possesses GUI (Graphical User Interface) reading capabilities. It can "view" a web page or a smartphone screen, identify buttons and input fields, and perform multi-step operations—such as planning a trip by navigating through multiple apps and eventually sending a summary email—without requiring the app developers to provide a custom interface. Slime: An Open-Source Reinforcement Learning Framework To support the training of such complex agents, Zhipu AI developed and open-sourced the Slime framework.[15] Slime addresses the bottleneck of reinforcement learning (RL) for agents by decoupling the training engine from the rollout engine.[15] This allows for the asynchronous generation of training data, meaning the GPUs used for training are never idle while waiting for the model to "explore" an environment.[15] This in infrastructure was instrumental in training the GLM-4.5 series to handle the high-uncertainty environments typical of agentic tasks. Commercialization and the MaaS Ecosystem While Zhipu AI maintained an academic heart, its commercial strategy was aggressively execution-oriented. The company pioneered the Model-as-a-Service (MaaS) model in the Chinese market, effectively becoming the "OpenAI + Palantir" of the region. BigModel.ai and Developer Ecosystem The BigModel.ai platform serves as the commercial gateway to Zhipu’s intellectual property. By providing standardized APIs for the GLM, Cog, and Code families, the platform has cultivated an ecosystem of over 2.7 million developers and 12,000 corporate clients. [4] Notably, Zhipu AI reported that 9 of the top 10 internet companies in China utilize its models, indicating its penetration into the highest tiers of the economy. Private Deployments and Sovereign AI A critical component of Zhipu’s commercial success is its “industry All-in-One” (Industry All-in-One) and private deployment services. For sectors such as finance, energy, and government—where data security and “sovereign control” are paramount—Zhipu provides the ability to train and deploy models within private cloud environments. This focus on the “thickness” of the model (depth of integration and security) has allowed Zhipu to secure high-value contracts that are often inaccessible to purely C-end focused competitors. The Financial Trajectory: From Eight Rounds of Funding to the Global IPO Zhipu AI’s journey to the public market was supported by a massive influx of capital from a diverse range of institutional and corporate investors. Over its six-year history prior to listing, the company completed eight rounds of financing, raising a cumulative total of more than 8.3 billion RMB. Strategic Capital Integration The investor list for Zhipu AI represents a cross-section of the global financial and technological landscape. It includes domestic internet giants (Alibaba, Tencent, Meituan, Xiaomi, Ant Group), prestigious venture capital firms (HongShan, Hillhouse, Qiming Venture Partners), and various state-owned investment vehicles. By the end of 2025, Zhipu AI’s valuation had surged to between 44 billion and 51.1 billion HKD. This financial strength allowed the company to maintain an R&D spend that reached 2.19 billion RMB in 2024, ensuring that it could continue to compete with the massive compute budgets of Western rivals. The Landmark IPO on the Hong Kong Stock Exchange On December 30, 2025, Zhipu AI officially launched its recruitment for an IPO on the Main Board of the Hong Kong Stock Exchange, under the stock code
- The offering price was set at
- 20 HKD per share, with the goal of raising approximately 4.3 billion HKD. he successful listing on January 8, 2026, was a historic event. It marked Zhipu AI as the first company globally to go public with a core business model built around AGI foundation models. This achieved two critical objectives:
- Capital Liquidity: It provided the necessary capital for the next phase of the "scaling law" race, specifically for the acquisition of compute and the expansion of the R&D team.
- Market Validation: It established a public-market benchmark for the valuation of large model companies, effectively beating Western leaders OpenAI and Anthropic to the public stage. Analysis of Global Competitive Positioning Zhipu AI’s strategy can be understood as a middle path between the pure-product focus of Western startups and the integrated industrial approach characteristic of Chinese tech giants. Zhipu AI vs. The "Six Tigers" of China Within the domestic "Six Tigers" (six leading AI startups), Zhipu AI is distinguished by its academic "thickness" and its B-end (enterprise) depth. While competitors like MiniMax have seen significant success in C-end applications—such as the “Talkie” social app which leads in North American markets—Zhipu has focused on building a “full-chain” infrastructure. Zhipu AI vs. OpenAI and Anthropic On the global stage, Zhipu AI is often referred to as “China’s OpenAI,” but its operational philosophy is distinct. While OpenAI has moved toward a more closed, product-centric model, Zhipu has maintained a robust commitment to open-source models (like ChatGLM and CogVideo) to seed the developer ecosystem. Technically, the GLM-4.5 series has demonstrated that Zhipu can match or exceed the performance of GPT-4.1 and Claude
- 5 Sonnet on critical benchmarks, particularly in mathematics and coding, while utilizing The decision to go public first is also a notable strategic divergence. While OpenAI and Anthropic have relied on near-infinite private funding rounds (e.g., Microsoft and Amazon/Google investments), Zhipu AI chose the public market to diversify its capital base and establish institutional legitimacy at a global scale. The evolution of Zhipu AI from an academic spin-off to a global public company is a narrative of strategic alignment. By combining the intellectual heritage of Tsinghua KEG with an aggressive commercial “Model-as-a-Service” strategy, the organization has created a blueprint for the sustainable development of AGI. Its success suggests that the next decade of AI development will not be defined solely by the size of a model’s parameters, but by its “Agentic” capabilities—the ability to act as an “Operator” within the human environment. As Zhipu AI transitions into its life as a public entity, the challenges will shift toward maintaining its R&D velocity while meeting the quarterly expectations of the Hong Kong Stock Exchange. However, with its "dual-driven" technical engine, its massive domestic market share, and its newly acquired public capital, the organization is uniquely positioned to lead the world into the era of the "Autonomous Intelligent Agent". Stock code 2513 is more than just a financial identifier; it is a signal that the era of laboratory-bound AGI is over, and the era of commercially integrated, action-oriented artificial intelligence has begun.