The landscape of generative artificial intelligence in China reached a historic milestone in January 2026 as two of the nation’s leading “AI tigers” – Beijing-based Zhipu AI and Shanghai-based MiniMax – successfully completed initial public offerings on the Hong Kong Stock Exchange (HKEX). These debuts not only secure the capital necessary for the global AI arms race but also highlight two distinct philosophies for achieving Artificial General Intelligence (AGI).
The divergence between these two organizations begins with their institutional heritage. Zhipu AI is a direct spin-off from the Knowledge Engineering Group (KEG) at Tsinghua University. Led by Professors Tang Jie and Li Juanzi, the company maintains a research-intensive DNA, prioritizing foundational AI infrastructure and knowledge engineering. Zhipu positions itself as the provider of a “national foundational platform,” focusing on on-premise and cloud-based deployments for state and enterprise sectors including finance, healthcare, and manufacturing.
In contrast, MiniMax represents the entrepreneurial spirit of the Shanghai technology cluster. Established by computer vision veterans from SenseTime, most notably founder Yan Junjie, the company adopts an industrial, product-first ethos. While Zhipu builds the bedrock, MiniMax focuses on the application layer, developing multimodal, consumer-facing products.
The technical strategies of both firms center on the Mixture of Experts (MoE) architecture, yet they prioritize different performance dimensions.
- Zhipu AI – GLM-4.7: This flagship model features 358 billion parameters (with 32 billion activated per token) and focuses on high-fidelity reasoning. A key innovation is its “Deep Thinking” mode, which enables multi-step planning and autonomous error recovery – capabilities that allow it to solve complex software engineering tasks with a 73.8% success rate on the SWE-bench Verified leaderboard. Its AutoGLM agent further reflects this “action-oriented” research, autonomously executing tasks on mobile devices via voice and visual prompts.
- MiniMax – M2.1: The company has targeted the computational constraints of traditional Transformers through a hybrid “Lightning Attention” architecture. This allows its M2.1 model to support an unprecedented native context window of 4 million tokens – roughly equivalent to 60 books. By enabling agents to “remember” entire codebases or massive document sets in a single prompt, MiniMax has pioneered a “RAG Killer” paradigm, potentially eliminating the need for complex Retrieval-Augmented Generation pipelines in many enterprise use cases. MiniMax’s products, including text-to-video generation (Hailuo AI) and multilingual speech models, cater directly to developers and end-users, emphasizing practicality and scalability.
Zhipu AI remains heavily domestic-focused after being added to the U.S. Entity List in January 2025. The designation restricts Zhipu’s access to high-end U.S. hardware and software, forcing it to focus on building a self-reliant domestic AI stack for Chinese institutional clients.
MiniMax, unencumbered by such restrictions, has pursued an aggressive international strategy. With nearly 73% of its revenue generated from overseas markets by late 2025, MiniMax has successfully engaged global developers through roadshows in cities like San Francisco and New York. Its consumer social app, Talkie, maintains over 32.8 million monthly active users, positioning MiniMax as a direct global competitor to Western labs like Character.AI.
As of 2026, Zhipu AI and MiniMax represent the two pillars of China’s generative AI ambition. Zhipu AI provides the technical and institutional “deep tech” foundation necessary for national sovereignty, while MiniMax demonstrates the country’s ability to produce world-class, cost-efficient consumer applications. Together, their simultaneous public debuts signal that China’s AI sector has moved past the experimental stage into a mature era of commercial validation and global competition.
