LatePost Exclusive | Zhipu Goes Public, Tang Jie’s Internal Letter Calls for Full Return to Foundational Model Research

The true determinants of the next stage's landscape still lie in two fundamental aspects—model architecture and learning paradigms. At the same time, a clear direction may emerge for applications: the breakthrough year for AI replacing different professions/tasks.

By丨Shen Yuan

Editor丨Song Wei

LatePost exclusively learned that on January 8, the day Zhipu was listed, Professor Tang Jie from the Department of Computer Science at Tsinghua University, founding initiator and Chief Scientist of Zhipu, issued an internal letter announcing that the next-generation model GLM-5 would soon be launched.

Tang Jie said, today's "an exciting day in Zhipu's life." He did not directly address the controversy over the business model of large model companies or state Zhipu's commercialization goals for 2026, but emphasized that the real achievement on the road to AGI for Zhipu is to have "real users" and to develop theories, technologies, or products that can help more people.

DeepSeek has created a shockwave among Chinese large model companies. Many believe that DeepSeek's phenomenal success first impacted Zhipu's ecological niche, as both share almost identical academic research team attributes, and Zhipu has also made significant contributions to the open-source ecosystem of large models.

The internal letter stated that Zhipu fulfilled its 2025 strategy as scheduled: launching a "stabilizing" model in April, a model that "got a seat at the table" mid-year (becoming one of the best), and releasing a Top 1 model by the end of the year.

This comprehensive return to foundational model research is Zhipu's response to the impact of DeepSeek. On December 23, Zhipu's base model GLM-4.7 was launched and open-sourced. Artificial Analysis (AA Intelligent Index) shows GLM-4.7 ranked first among domestic models and tied for sixth globally with Claude 4.5 Sonnet.

In addition to the release of GLM-5, the internal letter also introduced Zhipu's three technical focuses for 2026, including a brand new model architecture design, a more general RL (reinforcement learning) paradigm, and exploration into continuous learning and autonomous evolution of models. All these revolve around enhancing foundational model capabilities.

With the improvement of foundational model capabilities, Agents and domain models will ultimately combine with the foundational models. In fact, AI may not even necessarily require the creation of new applications. "The application of large models must also return to first principles." As Tang Jie said in a Weibo post last year, 2026 will be the breakout year for AI to replace various professions.

In 2025, Zhipu also underwent substantial organizational restructuring, reducing the scale of To C, product-development, and video-generation teams. Achievements including AutoGLM have been open-sourced one after another.

Since the release of ChatGPT, more than three years of rapid AI development have passed, "There is no real consensus in the industry, everyone is just moving forward." Tang Jie mentioned in an internal discussion.

Below is the full text of Tang Jie's open letter, exclusively released by LatePost with Zhipu's authorization.

Doing AGI with a "coffee" spirit

During a short visit to the Hong Kong University of Science and Technology, I ran into Professor Yang Qiang at the café on the first floor of the lab. I said I'd been drinking too much coffee these days, feeling a bit addicted, and should probably cut back.

Professor Yang said, "Why quit? Addiction isn't necessarily a bad thing. If we could be as addicted to research as we are to coffee, wouldn't we do great research?"

Indeed, "addiction" is what makes life exciting. Whether in research or other things, as long as we focus and work hard, we can achieve great things.

"Letting machines think like humans" has always been Zhipu's vision and ideal, as well as the sole goal Zhipu people persistently strive for.

At the end of 2018, inspired by the dual-system theory of human cognition, we designed a machine "cognitive" system incorporating fast and slow thinking. In 2019, we officially founded Zhipu and began exploring AGI, aiming to realize the grand vision of "letting machines think like humans."

The greatest challenge here, perhaps, is that even to this day, no one—including ourselves—has been able to give an accurate definition of AGI or a clear technical path to achieve it. Maybe that's the charm of exploring AGI.

We are at an unprecedented and extraordinary historical moment—a time when technology is once again revolutionizing the world. Large models are not only the key foundation for general artificial intelligence but are also poised to become the core engine driving productivity transformation.

Looking back, a crucial reason we have come this far is that we have always insisted on developing AI technologies that users can truly use. Only theories, technologies, or products that have real users can become significant milestones on the road to AGI. Of course, not all innovations succeed; we have had many risky projects that ended in failure. Yet, these failures taught us to draw strength from setbacks, making Zhipu stronger and deepening our understanding of AGI. More importantly, this focus on practicality made us look beyond short-term gains: helping users, the country, and advancing global technology have become Zhipu's long-term goals.

In 2020, we launched our own large model algorithm architecture GLM and began training a base model with 10 billion parameters. The model was successful and was trialed by many companies, including Meituan. This was a bold attempt since it was still the era of small BERT models. But the success at that time was still far from our AGI dream. Partly because the model's knowledge base was not large enough; partly because it couldn't yet reason and think like a human.

From 2021 to 2022, the development of large models was not smooth. Most people did not accept the "let machines think like humans" vision, considering it as far-fetched as going to the moon, nor did they see it as a major technological disruption—or they were simply afraid of failure. We still decided to gamble and trained a large model with 130 billion parameters using even more data.

This decision was difficult because it could not disrupt the company's overall pace of development. We thus set up two dedicated small innovation teams: one focused on model training (which became the GLM trio), and the other independently built the MaaS platform—at that time, the two teams might not even have known of each other's existence. By mid-2022, GLM-130B was trained, with many fine designs drawing global attention; meanwhile, the MaaS platform (now bigmodel.cn) went live and got its first batch of real API users. Later, we officially established the AI Institute in the company, dedicated to next-generation large model R&D, and formed the MaaS Platform Department to provide large model API services externally. Sometimes, we need to find people with bold dreams (even devoting extra effort to finding such people), as a bold and grand goal may determine half the success.

In 2023, I discussed with a top domestic entrepreneurial peer (actually much younger than me) about AI's potential to reshape the future. We both believed AI would disrupt search, disrupt browsers, and bring every person a brand new AI assistant. Once we have this AI assistant, we might not need app stores anymore; instead, we may need to build an "API store" for AI, and the underlying logic of this API store could disrupt existing operating systems. Perhaps, an even more profound disruption would be to computers themselves, as we might then need computers tailored for AI rather than for humans.

The significance of this revolution will be immense, as it will completely reshape the underlying logic of computers, challenging the von Neumann architecture—the cornerstone of computing for 80 years. At this point, we both felt that our investment in AI was still too little, not "all-in" enough.

Reality is also harsh—going "all-in" requires not only strong conviction but also immense financial and team support, as well as accurate foresight. 2023–2024 is the breakout period for global large models; major companies are "all-in" on large models, and a wave of entrepreneurship has swept across China, with a "hundred-model war" and various AI assistants emerging in rapid succession.

We made mistakes too, both technical and commercial. Looking back, perhaps the reason is that we sometimes lost our way on the road to AGI, distracted by short-term gains and temporary excitement. AGI is a technological revolution—technology should be equitable, open, transparent, and benefit everyone.

The emergence of DeepSeek woke us up. When Wenfeng started his business in 2023, we talked, but I didn't realize how dedicated he was to AGI. I thank him for giving me many new perspectives. The choice to relentlessly pursue AGI technology, to keep exploring the upper limits of AGI, and to make precise future predictions are areas where Zhipu must continuously improve and elevate itself. These past two years have taught us a lot, especially "reinforced" our understanding of AGI, company governance, and business competition.

In the past year, we actually carried out a systematic "reinforcement." We called for "perseverance" and "achievement," asking everyone to remain steadfast, neither arrogant nor self-deprecating, to accomplish their own tasks and realize themselves.

At the beginning of the year, everything was so difficult—the model's effects didn't meet expectations, there was a price war nationwide, and breaking through required finding a precise entry point.

We held our ground and finally identified coding as our breakthrough point.

If the release of GLM-4.1 in April was a symbolic probe, then the launch of GLM-4.5 at the end of July was practically a decisive battle. All technical, platform, and business teams were on edge, working tirelessly day and night. Eventually, we achieved a long-awaited victory. Then, GLM-4.6 and GLM-4.7 enabled our models to stand shoulder to shoulder with top international models. Our GLM-4.7 achieved SOTA among open-source and domestic models in multiple evaluations, including AA and Arena. The real-world coding and Agent experiences from users were also excellent. Developers from 184 countries around the world—over 150,000—used the GLM Coding Plan. After the release of GLM-4.7, the annualized revenue of the MaaS platform exceeded 500 million (with overseas revenue exceeding 200 million), growing from 20 million to 500 million (a 25-fold increase) in just 10 months.

Overall, on the model side, we completed our overall strategy as scheduled: a "stabilizing" model in April, a "seat at the table" model (one of the best) mid-year, and a Top 1 model by the end of the year. This laid a solid foundation for our continued pursuit of AGI's technological frontier.

Our "Sovereign AI" has also made significant progress: Malaysia's national MaaS platform was built on the Z.ai open-source model, making GLM a national model in Malaysia. The overseas expansion of Sovereign AI was inspired by the General Secretary's call at a symposium for "China's AI to go global." To be honest, I didn't know how to proceed at first, but our international team was daring and determined, achieving a milestone for China's large models in going overseas. On the business front, we dared to compete and achieved annual revenue growth that once again exceeded double the target.

Amidst all the challenges and opportunities, today, we have become the world's first listed large model company in an almost impossible way, showing the market's recognition of our technological and commercial value. "Make impossible possible"—do you remember when we said this?

Over the past year, perhaps the biggest transformation was not Zhipu itself, but a group of young people on the front lines who made the impossible possible.

In 2026, our goal is to become an internationally leading large model enterprise. In the past year, many discussions about large models have focused on applications and ecosystems.

The true determinants of the next stage's landscape still lie in two fundamental aspects—model architecture and learning paradigms. At the same time, a clear direction may emerge for applications: the breakthrough year for AI replacing different professions/tasks.

Based on this judgment, our focus for 2026 will be:

• GLM-5. GLM-5 will soon meet everyone. Through further scaling and many technological improvements, we believe GLM-5 will bring many novel experiences and help AI complete more real-world tasks for everyone.

• A brand new model architecture design. The Transformer architecture, widely used for nearly 10 years, is showing some shortcomings, including computation overhead for super long contexts, memory mechanisms, update mechanisms, etc. All these require exploring new model architectures, discovering new scaling paradigms, and improving computational efficiency via chip-algorithm co-design and other technologies.

• RL with stronger generalization ability. The current mainstream RLVR paradigm, despite its success in mathematics and code, increasingly shows limitations due to its reliance on artificially constructed verifiable environments. This year, we need to explore more general RL paradigms, enabling AI not only to complete specific tasks under human instructions but also to understand and execute long-term tasks spanning hours or even days.

• The most challenging exploration is opening the path to continuous learning and autonomous evolution. Currently, all mainstream AI models possess intelligence that is essentially static after deployment. They acquire knowledge through a one-time, costly training process and then gradually become outdated in application. This is in stark contrast to the human brain, which can continuously learn and evolve through interaction with the world. We need to make forward-looking plans for the next-generation learning paradigm—online learning or continual learning.

We are not a traditional company, nor do we intend to become one. We hope to be an AI-native company where anything is possible: building next-generation models that continually push the limits of intelligence, developing AI-centered products and services for users. We want AI to become everyone's most capable assistant, helping us accomplish our tasks. We also believe we should use AI to participate in company governance, reduce costs, improve efficiency, and achieve greater fairness.

As time goes on, a company often gets used to doing the same things, making only incremental improvements, which limits our innovation. But in the age of AI, everything is revolutionary. We need to feel a little "uncomfortable" to keep innovating and to propose revolutionary ideas that drive the next big area of growth.

Therefore, we have established a brand-new department within Zhipu called X-Lab, dedicated to gathering more young people through open means and conducting cutting-edge explorations, including new model architectures, new cognitive paradigms, and incubating new projects—not limited to software or hardware. At the same time, we will expand our external investments, not just strategically cooperating with existing invested companies, but also opening up new territories, so the entire industry and ecosystem can prosper together. At X-Lab, everyone's mission is to achieve truly disruptive innovation, ultimately returning to the main thread of AGI.

Today is an exciting day in Zhipu's life, an important milestone in Zhipu's history, and the start of a brand-new era for Zhipu. I really like the brand Z.ai—Z is the last letter of the alphabet, representing the ultimate state. We hope to reach the ultimate state of intelligence on the journey to explore AGI, and that is our goal. We are very excited:

- To have an ambitious, world-changing cause

- To focus on long-term interests and look to the future

- To be more focused and explore the essence of AGI

- To empower great entrepreneurs and companies with AI for their vigorous development

- To seize development opportunities for enterprises with more accurate foresight

- Ultimately, we hope to bring a different kind of AI to human society and truly promote human well-being.

This is an unparalleled moment of joy, not the fleeting dopamine of the moment, but the endorphins accumulated on the path to AGI, making us more focused, grounded, and persistent in moving forward!

Tang Jie

2026.1.8

Header image source: "Dune 2"

- FIN -