Zhipu in the stock market: Tang Jie outlines the path to AGI, focusing on foundational models and new architectures

The listing of Zhipu marks a crucial moment in the global artificial intelligence landscape. In his speech at the January 8th listing event, Tang Jie, the company’s founder and a professor at Tsinghua University, outlined an ambitious vision for the company’s next chapter: a focused return to pure research on foundational models, abandoning short-term commercial distractions to pursue Artificial General Intelligence.

Achievements in 2025: From Strategy to Reality

Zhipu has stayed true to the commitments announced at the beginning of the year. The roadmap outlined three well-defined phases: in spring, the launch of a model capable of “defending its position”; mid-year, the release of a “high-end” model capable of competing at the highest levels; and towards the end of the year, the debut of a Top 1 performance model. This strategy proved successful.

The decisive turning point came with GLM-4.5 in July, when all teams worked in sync to achieve a qualitative leap. Subsequently, the releases of GLM-4.6 and GLM-4.7 solidified Zhipu’s competitive position among Chinese open-source models. According to Artificial Analysis, GLM-4.7 ranks first among Chinese models and sixth globally, only comparable to Claude 4.5 Sonnet.

On the commercial front, the MaaS platform experienced exponential growth: annualized revenue exceeded $500 million after the launch of GLM-4.7, with over $200 million coming from international markets. In just 10 months, the platform grew from 20 to 500 million (a 25-fold increase). The GLM Coding Plan has over 150,000 developers from 184 countries.

The Challenge of DeepSeek and Returning to Fundamentals

The emergence of DeepSeek has served as an important wake-up call for the industry. Although the startup attracted market attention, the event prompted Zhipu to reconsider its priorities: during periods of rapid AI development, many companies dispersed into vertical applications, niche AI assistants, and short-term commercial strategies. Tang Jie acknowledges that Zhipu itself made mistakes during the “battle of a hundred models” between 2023 and 2024.

The lesson learned is clear: AGI is a technological revolution that requires determination and long-term vision. Technology must be accessible to all and bring widespread benefits, not focus on fleeting profits. For this reason, Zhipu’s focus in 2026 will be a complete return to pure innovation in foundational models.

The Vision for 2026: Three Technological Pillars

Throughout 2026, Zhipu will focus its efforts on three strategic directions that will shape the next phase of AI:

GLM-5 and innovative scaling: The new next-generation model will soon be available to everyone. Thanks to further scaling technologies and architectural innovations, it is expected to offer significantly new experiences and assist users in completing more complex real-world tasks.

Revolutionary architectures: The Transformer architecture, dominant for nearly a decade, has obvious limitations in computational costs for large contexts, memory mechanisms, and updates. Zhipu aims to discover new scaling paradigms, explore alternative architectures, and implement chip-algorithm co-design to improve computational efficiency.

Generalized RL paradigms: The currently dominant RLVR approach has shown success in mathematics and programming, but its dependence on manually verified environments limits its universal applicability. Next year, the company intends to develop more generalized RL paradigms, enabling AI not only to perform tasks on instruction but to understand and complete long-term tasks lasting hours to days.

The Frontier of Continuous Learning

The most ambitious challenge concerns continuous learning and autonomous evolution of models. Current AI systems remain static after deployment, learning through an expensive single training process before becoming obsolete. The human brain, by contrast, learns and evolves constantly through interaction with the environment. Zhipu will proactively plan the next-generation learning paradigm—Online Learning or Continual Learning—representing the next big leap toward AGI.

Governance, Internationalization, and the New X-Lab

Zhipu does not intend to become a traditional company. It has internally established a new department, X-Lab, dedicated to gathering young talents with an exploratory spirit for cutting-edge research on new architectures, cognitive paradigms, and incubation of innovative projects.

On the international front, the “Sovereign AI” initiative has made significant progress: Malaysia’s national MaaS platform is built on the open-source Z.ai model, positioning GLM as Malaysia’s national model. This marks the first successful attempt to bring Chinese large models into the global market.

For 2026, the company’s goal is to become an international leader in large models, while maintaining a commitment to authentic AGI. As Tang Jie states, the true measure of success is not in commercialization goals, but in having “real users” and developing theories, technologies, or products that can truly help more people in the global scientific research progress.