What will truly determine the competitive landscape in the next phase are two more fundamental factors: model architecture and learning paradigms. Concurrently, a clear direction may emerge on the application side: a breakout year for AI replacing various job roles and tasks.
It was exclusively learned that on January 8th, the day of Knowledge Atlas's listing, Tang Jie, a professor in Tsinghua University's Computer Science Department and the founding initiator and Chief Scientist of Knowledge Atlas, released an internal letter announcing the upcoming launch of the new-generation model GLM-5.
Tang Jie described the day as "an exciting day in the life of Knowledge Atlas." He did not directly address the controversies surrounding the business models of large model companies or specify Knowledge Atlas's 2026 commercialization targets, but he emphasized that theories, technologies, or products that are genuinely "used by people" and can help more individuals represent the significant achievements for Knowledge Atlas on its path toward AGI.
The rise of DeepSeek has sent shockwaves through China's large model enterprises. Many believe that DeepSeek's phenomenal success has first impacted Knowledge Atlas's niche, as both share similar academic and research team attributes, and Knowledge Atlas has also made substantial contributions to the open-source ecosystem for large models.
The internal letter stated that Knowledge Atlas successfully completed its strategy set at the beginning of 2025: releasing a model in April to "hold its ground," a model by mid-year to "get a seat at the table" (becoming one of the best), and a Top 1 model by year-end.
This strategy of fully returning to foundational model research is Knowledge Atlas's response to the impact of DeepSeek. On December 23rd, Knowledge Atlas launched and open-sourced its foundational model GLM-4.7. According to the Artificial Analysis (AA) index, GLM-4.7 ranks first domestically and is tied for sixth globally with Claude 4.5 Sonnet.
Beyond the announcement of GLM-5, the internal letter also outlined three technical directions Knowledge Atlas will focus on in 2026: entirely new model architecture design, more generalized RL (Reinforcement Learning) paradigms, and exploration into models' continuous learning and autonomous evolution. All these directions revolve around enhancing foundational model capabilities.
As foundational model capabilities improve, Agents and domain-specific large models will ultimately integrate with foundational models. Furthermore, AI does not necessarily mean creating entirely new applications. "The application of large models must also return to first principles," Tang Jie said in a Weibo post last year, predicting that 2026 will be a breakout year for AI replacing different job roles.
In 2025, Knowledge Atlas also underwent significant organizational adjustments, scaling back its To C, product R&D, and video generation teams, while open-sourcing achievements including AutoGLM.
Since the release of ChatGPT, over three years of rapid AI development, "the industry hasn't reached much consensus; everyone is just moving forward," Tang Jie remarked during an internal exchange.
The following is the full text of Tang Jie's open letter, exclusively authorized by LatePost for publication.
Embodying the 'Coffee' Spirit in Pursuing AGI
During a short-term visit to the Hong Kong University of Science and Technology, I happened to run into Professor Yang Qiang at the coffee shop on the first floor of the lab. I mentioned that I had been drinking too much coffee lately, feeling somewhat addicted, and needed to cut back.
Professor Yang said, "Why quit? Addiction isn't necessarily a bad thing. If we could become as addicted to research as we are to coffee, why worry about not doing research well?"
Indeed, 'addiction' is what makes life exciting. Whether it's research or anything else, with focus and effort, we can certainly excel.
"Enabling machines to think like humans" has been the consistent vision and ideal of Knowledge Atlas, and the sole goal for which its people persistently strive.
Inspired by the dual-system theory of human brain cognition, we designed a machine 'cognitive' system featuring fast thinking + slow thinking around the end of 2018. In 2019, we formally established Knowledge Atlas and began exploring AGI, aiming to realize the grand vision of "making machines think like people."
The greatest challenge here might be that even today, no one, including ourselves, can provide a precise definition of AGI or a clear technical path to achieve it. Perhaps this is precisely the allure of exploring AGI.
We are at an extraordinary moment in history, a time when technology is once again disruptively changing the world. Large models are not only the critical foundation for artificial general intelligence but also hold the potential to become the core engine driving productivity transformation.
Looking back on our journey, a key reason we have reached this point is our unwavering commitment to developing AI technologies that users can genuinely utilize. Only theories, technologies, or products that are truly used by people can ultimately become significant achievements on the path to AGI. Of course, not all innovations succeed; we have had many risky projects that ended in failure. But these failures have often taught us to draw strength from setbacks, making Knowledge Atlas stronger and deepening our understanding of AGI. More importantly, this has led us to focus on practicality while no longer fixating solely on short-term gains: helping users, contributing to the nation, and advancing global technological progress have become Knowledge Atlas's long-term goals.
In 2020, we introduced our own large model algorithm architecture, GLM, and began experimenting with training a foundational model with 10 billion parameters. The model was quite successful, receiving trial use from many enterprises, including Meituan. This was a bold attempt because it was still the era of small models dominated by BERT. However, that success was still far from the AGI of our dreams, partly because the model's knowledge capacity was insufficient, and partly because it couldn't reason and think like a human.
From 2021 to 2022, the development of large models was not smooth. Most people did not accept the "making machines think like people" plan, viewing it as a moon-shot-like crazy endeavor. They either didn't see it as a major technological transformation opportunity or feared failure. We decided to take a gamble, training a large model with 130 billion parameters using more data.
This decision was difficult because it couldn't disrupt the company's overall development rhythm. To manage this, we established two dedicated small innovation teams: one responsible for model training, which later became the company's "GLM Three Musketeers"; the other independently responsible for building the MaaS platform. At the time, these two small teams might not have even known of each other's existence. By mid-2022, GLM-130B was trained, with many of its intricate designs attracting global attention. Simultaneously, the MaaS platform launched (now bigmodel.cn), gaining its first real API users. Subsequently, we formally established the AI Institute within the company to focus on next-generation large model R&D, and the MaaS Platform Department to provide external large model API services. Sometimes, we need to find people bold enough to dream big (even dedicating extra effort to find such individuals), as a bold, grand goal can determine half the success.
In 2023, I discussed the potential changes AI could bring with a top domestic entrepreneurial veteran (who is actually younger than me). We agreed that AI would颠覆 search,颠覆 browsers, and bring a brand-new AI assistant to everyone. Furthermore, with this AI assistant, we might no longer need app stores; instead, we might need to create an "API store" for AI, potentially颠覆 the underlying logic of existing operating systems. Subsequently, an even greater disruption might be the computer itself, as we might not need computers designed for humans anymore, but rather computers suited for AI.
The significance of this transformation would be immense, as it would fundamentally reshape the underlying logic of computing, challenging the 80-year-old cornerstone of computer science—the von Neumann architecture. Reaching this point in the conversation, we simultaneously felt that our investment in AI was still too little, not yet "All-in."
Reality is also harsh. Going "All-in" requires not only firm belief but also exceptionally strong financial and team support, coupled with precise foresight. The years 2023 to 2024 were the breakout years for large models globally. Major tech firms纷纷 declared they were "All-in" on large models, a wave of startups surged domestically, leading to the "Hundred Model Battle," and various AI assistants emerged one after another.
We likely made mistakes during that time, both technical and commercial. In retrospect, the reason might be that we occasionally lost our way on the path to AGI, distracted by short-term gains and the hype of the moment. AGI is a technological transformation; technology is egalitarian, open, and transparent, meant to be accessible and beneficial to all.
The subsequent emergence of DeepSeek served as a wake-up call for us. Wen Feng spoke with me when he started his venture in 2023; at the time, I didn't realize his profound dedication to AGI. I thank him for bringing many different perspectives. Choosing a persistent pursuit of AGI technology, continuously exploring the upper limits of AGI, and making accurate future predictions have become areas where Knowledge Atlas needs constant improvement and升华. The past two years have taught us a lot; more importantly, we have undergone "reinforcement learning" regarding our understanding of AGI, corporate governance, and商业 competition.
Over the past year, we have actually conducted a systematic "reinforcement." We championed the slogans of "Perseverance" and "Achievement," urging everyone to maintain perseverance, avoid arrogance or rashness, not undervalue themselves, achieve success in the tasks at hand for Knowledge Atlas, and achieve personal growth.
At the beginning of the year, everything was incredibly difficult. Model performance fell short of expectations, and a nationwide price war made breaking through require finding a precise突破口.
We held our ground and finally identified Coding as our突破口.
If the release of GLM-4.1 in April felt like a symbolic probe, the launch of GLM-4.5 at the end of July was almost a decisive battle. All technical, platform, and business teams held their breath, working overtime day and night. We finally welcomed a long-awaited victory. Subsequently, GLM-4.6 and GLM-4.7 allowed our model capabilities to catch up with those top international models. Our GLM-4.7 achieved SOTA status for open-source models and domestic models in multiple evaluations, including AA and the Arena. User feedback on model Coding and Agent performance has also been very positive. 150,000 developer friends from 184 countries use the GLM Coding Plan. Following the release of GLM-4.7, the MaaS platform's Annual Recurring Revenue (ARR) exceeded 500 million RMB (with overseas revenue exceeding 200 million RMB), growing 25-fold from 20 million to 500 million in just 10 months.
Overall, on the model side, we successfully completed the annual strategy set at the beginning of the year: releasing a model in April to "hold its ground," a model by mid-year to "get a seat at the table" (becoming one of the best), and a Top 1 model by year-end. This has laid a crucial foundation for our continued冲刺 towards the high ground of AGI technology.
Our 'Sovereign AI' initiatives also made significant progress: Malaysia's national-level MaaS platform was built based on Z.ai's open-source models, making GLM a national-level model in Malaysia. The Sovereign AI出海 initiative was inspired by my participation in a symposium with the General Secretary and his call for "Chinese AI to go global." Honestly, I wasn't entirely sure how to proceed, but our international team dared to fight and achieve, accomplishing a milestone in taking Chinese large models abroad from scratch. On the business front, we dared to compete and once again achieved our goal of over 100% annual revenue growth.
Amidst various challenges and opportunities, today, we have achieved the nearly impossible feat of becoming the world's first listed large model company. This demonstrates market recognition of our technological and商业 value. "Make impossible possible"—remember what we used to say?
Over the past year, perhaps the greatest transformation hasn't been within Knowledge Atlas itself, but among a group of young frontline employees who have truly accomplished many things that seemed impossible.
The company's goal for 2026 is to become an internationally leading large model enterprise. Over the past year, much of the discussion about large models has centered on applications and ecosystems.
What will truly determine the competitive landscape in the next phase are two more fundamental factors: model architecture and learning paradigms. Concurrently, a clear direction may emerge on the application side: a breakout year for AI replacing various job roles and tasks.
Based on this assessment, in 2026 we will focus on:
GLM-5. GLM-5 will be released soon. Through further Scaling and numerous entirely new technical improvements, we believe GLM-5 will bring many novel experiences, enabling AI to help everyone complete more real-world tasks.
Entirely new model architecture design. The Transformer architecture, widely used for nearly a decade, is showing limitations, including computational costs for ultra-long contexts, memory mechanisms, and update mechanisms. This necessitates exploring entirely new model architectures, discovering new Scaling paradigms, and improving computational efficiency through techniques like chip-algorithm co-design.
RL with stronger generalization capabilities. While the current mainstream RLVR paradigm has succeeded in mathematics and coding, its limitations, relying on manually constructed verifiable environments, are becoming increasingly apparent. This year requires exploring more generalized RL paradigms, supporting AI not only to complete specific tasks under human instruction but also to understand and execute long-duration tasks spanning hours or even days.
The most challenging exploration is to embark on the path towards continuous learning and autonomous evolution. The intelligence of all current mainstream AI models is essentially static after deployment. They acquire knowledge through a one-time, costly training process and gradually become obsolete in application. This is fundamentally different from the human brain's ability to continuously learn and evolve from ongoing interaction with the world. We need to make前瞻性布局 for the next generation of learning paradigms—Online Learning or Continual Learning.
We are not a traditional company, nor do we intend to become one. We aspire to be an AI-native company where anything is possible: developing next-generation models that continuously push the upper limits of intelligence, and creating products and services with AI at their core to serve users. We want AI to become everyone's most capable assistant, helping us complete tasks. We also believe in utilizing AI to participate in corporate governance for cost reduction, efficiency improvement, and achieving greater fairness.
Over time, a company often becomes accustomed to doing the same things, making incremental improvements, which can limit our创新能力. But in the AI era, everything is transformative. We need to feel a bit "uncomfortable" to maintain创新能力 and propose revolutionary ideas that drive the next major growth area.
Therefore, we have established a全新的 department within Knowledge Atlas called X-Lab. This department will be dedicated to gathering more young people in an open manner to conduct frontier exploration, including entirely new model architectures and new cognitive paradigms. It will also incubate new projects, not limited to software or hardware. Simultaneously, we will expand external investments, not only strategically collaborating with existing portfolio companies but also opening up new territories, interconnecting the entire industry, and fostering collective prosperity across the ecosystem. At X-Lab, everyone's mission is to pursue completely disruptive innovation, ultimately returning to the main thread of AGI.
Today is an exciting day in the life of Knowledge Atlas, a significant milestone in its history, and the beginning of a全新的 era for the company. I am very fond of the Z.ai brand. 'Z' is the last letter in the alphabet, representing the ultimate destination. We hope to reach the ultimate destination of intelligence on the journey of exploring AGI; this is the goal we strive for. We are tremendously excited: - To have an ambitious, world-changing cause - To focus on long-term interests and look towards the future - To be more focused, exploring the essence of AGI - To empower the flourishing development of great entrepreneurs and companies with AI - To seize enterprise development opportunities with more precise foresight - Ultimately, to hope we can bring a different kind of AI to human society, tangibly promoting the advancement of human welfare.
This is an unparalleled moment of joy. This joy is not a fleeting dopamine rush but the endorphins accumulated on the path to exploring AGI, making us more focused, keeping our feet on the ground, and propelling us forward continuously!