Applying reasoning to self-developed Agents, Zhipu has started to take the lead

Author | Liu Baodan

Editor | Huang Yu

DeepSeek has become an industry benchmark, and in this AI competition, mainstream model companies are trying to surpass DeepSeek, while the domestic AI unicorn Zhipu has also provided its own answer.

On March 31, Zhipu officially launched AutoGLM Reflection at the Zhongguancun Forum. This new intelligent agent not only possesses deep research capabilities but also enables practical operations, truly pushing AI Agents into the stage of "thinking while doing."

As the world's first agent that integrates deep research and practical operation capabilities, the release of AutoGLM Reflection marks an important advancement in Zhipu's autonomous intelligent agent technology and a further upgrade of device control agents.

Behind AutoGLM Reflection, Zhipu has launched the Agentic GLM series matrix, including the GLM-4 base model, GLM-Z1 inference model, GLM-Z1-Rumination Reflection model, and AutoGLM model, especially the inference model GLM-Z1-Air, which has inference performance comparable to DeepSeek R1, but at only 1/30 of the price of R1.

It has been over two years since ChatGPT gained widespread attention, and AI large models have shifted from technological iteration to practical application, the latter becoming the core indicator for assessing the competitiveness of model vendors. Currently, Zhipu has partnered with leading companies such as Samsung and has become a partner in cities like Beijing and Shanghai. In addition, Zhipu has also initiated an overseas expansion strategy.

Zhipu CEO Zhang Peng told Wall Street Journal that the company's commercial revenue is expected to grow by over 100% in 2024, with even greater growth opportunities in 2025.

In this AI large model competition, Zhipu is beginning to explore its own growth path.

AutoGLM Reflection: Thinking While Doing

Four months ago, Zhipu used AutoGLM to implement a group red envelope feature, marking the first red envelope sent by AI, representing the transition of AI large models from dialogue to operation. Four months later, Zhipu has laid out its inference capabilities onto Agents.

Unlike the last time when red envelopes were sent, this time Zhipu has started to make money with Agents.

Fourteen days ago, Zhipu secretly conducted a test, registering a Xiaohongshu account focused on lifestyle science popularization. Then, it used AutoGLM Reflection to generate notes, such as how to choose a coffee pot and how to compare cosmetic ingredients.

During the live demonstration, Zhang Peng stated that they investigated the three most popular anti-aging ingredients in cosmetics for 2025, comparing their effects, usage, advantages, and disadvantages, and finally conducting rigorous comparative analysis. "Some tasks are really not simple."

Zhipu revealed the account's achievements on-site: in two weeks, it gained 5,000 followers and received multiple business order invitations. Yesterday, Zhipu issued its first business order, earning 500 yuan.

AutoGLM Reflection is the first agent that combines deep research capabilities and web operation capabilities, reflecting Zhipu's latest understanding of AI Agents, which is to enable machines not only to think but also to take proactive actions, achieving the goal of "thinking while doing." This is also where AutoGLM's "Chensi" differs from OpenAI's "DeepResearch." "Chensi" evolves AI agents from mere thinkers into intelligent executors capable of delivering results.

Zhang Peng stated that "Chensi" has broken through real-time online search, dynamic tool invocation, deep analysis, and self-verification, achieving true long-range reasoning and task execution.

For example, asking "Chensi" to write a report on "how the success of Nezha 2's box office will change the Chinese film industry." According to Zhang Peng, the "Chensi" model excels at such open-ended questions that require the model to explore on its own, ultimately generating a report of nearly 10,000 words.

This time, Zhiyu released the preview version of AutoGLM "Chensi," which primarily supports research scenarios. Zhang Peng revealed that in the next two weeks, they will further expand the execution capabilities of more agents.

In addition, AutoGLM "Chensi" has been launched and is currently available on the Zhiyu Qingyan PC client.

Fully Self-Developed Model Stack Behind

Behind the AutoGLM "Chensi" model is Zhiyu's independently developed full-stack large model technology.

Overall, "Chensi" integrates the general capabilities of GLM-4, the reflective capabilities of GLM-Z1, the rumination capabilities of GLM-Z1-Rumination, and the automatic execution capabilities of AutoGLM.

Zhiyu retrained a base model, GLM-4-Air-0414, with 32 billion parameters, incorporating more code-related and reasoning-related data during the pre-training phase, and optimizing for agent capabilities during the alignment phase, significantly enhancing the model's abilities in tool invocation, online search, and other agent tasks.

At the conference, Zhang Peng stated that GLM-4-Air-0414, with its 32 billion parameters, is comparable to larger parameter models from both domestic and international mainstream sources, making it particularly effective for adapting to agent tasks. "This is because agent tasks often involve multiple rounds of complex interactions, and the 32 billion parameters allow GLM-4-Air-0414 to quickly execute complex tasks, providing a solid foundation for the large-scale application of AI agents."

Based on GLM-4-Air-0414, Zhiyu launched a new deep thinking model, GLM-Z1-Air, which can compete with DeepSeek-R1 (671B, activated 37B) in performance.

In terms of inference speed, GLM-Z1-Air is 8 times faster than R1, with costs reduced to 1/30, achieving a dual breakthrough in high performance and high cost-effectiveness. Additionally, GLM-Z1-Air can run on consumer-grade graphics cards.

Based on GLM-Z1, Zhiyu enhanced the model's long-range reasoning capabilities by expanding reinforcement learning training, resulting in the "Chensi" model GLM-Z1-Rumination.

Zhang Peng stated that this model breaks through the limitations of traditional AI, which relies solely on internal knowledge reasoning, innovatively combining real-time online search, dynamic tool invocation, deep analysis, and self-verification to form a complete autonomous research process GLM-Z1-Rumination can actively understand user needs, continuously optimize reasoning, and repeatedly verify and correct hypotheses in complex tasks, making research results more reliable and practical. Compared to traditional reasoning models, Zhipu is looking forward to the thinking model leading AI assistants into a stage of "high IQ" to "high IQ + high autonomy."

The core of the Agent is reasoning planning and hands-on ability. If the thinking model is the brain of AutoGLM's contemplation, then AutoGLM is the hands and feet of AutoGLM's contemplation.

Zhipu released AutoGLM last October, which is the world's first large model agent capable of executing over 50 steps of action on a mobile phone. The AutoGLM contemplation version has seen significant evolution compared to the last release.

Zhang Peng stated that there is a Scaling Law in the pre-training and post-training of large models, and Agents also have a similar Scaling Law. "Based on the Agent Scaling Law, we further discovered the emergence of capabilities in Agents."

For example, during training, AutoGLM contemplation was never taught to access the Giant Tide Information Network. However, when given the instruction "Help me collect yesterday's research reports on embodied intelligence," AutoGLM contemplation was able to plan a solution to the problem by accessing Giant Tide Information and successfully operated the website.

Zhang Peng stated that AutoGLM's hands-on ability is currently at the Sota level in the industry, with comprehensive leading capabilities in using tools including browsers, mobile phones, and computers. In terms of GUI agents, CogAgent has achieved Sota results on multiple rankings for GUI Agents.

The series of achievements mentioned above is inseparable from Zhipu's forward-looking layout for Agents.

From the earliest launch of Zhipu Qingyan with FunctionCall capability in October 2023, to the launch of GLMs supporting agent orchestration in January 2024, and then the release of AutoGLM in October 2024, and today’s launch of AutoGLM contemplation, Zhipu has been leading the exploration of Agents.

After six years of technological accumulation, Zhipu has finally begun to show more competitiveness in this AI competition.

Open Source Does Not Mean Free

Like AI companies such as DeepSeek and Alibaba, Zhipu also adheres to an open-source strategy. Zhang Peng stated that the aforementioned models will be open-sourced on April 14 and will be successively launched on the MaaS platform within the next two weeks.

The commercial logic behind Alibaba's open source is to sell cloud computing services. For model vendors, open source means making core technologies public, which poses certain challenges for commercialization.

Zhang Peng also candidly acknowledged that open source will have some impact on the commercialization market. However, he emphasized that open source does not mean completely free; the subsequent investment of technical personnel, maintenance costs, and how to localize DeepSeek will incur significant costs, requiring a professional team to solve the problems.

Currently, there are various agent products on the market. As a model vendor that started research on agents early, Zhipu is well aware of market demands. Zhang Peng emphasized: "We must provide services for the model; it's not enough to just throw the product at the enterprise. If the enterprise can't handle it, it's pointless. If they spend money and find it unusable, they will come back and tell you it's not good." Currently, Zhipu is vigorously promoting the overall service of AI technology, including providing tools and platforms, offering cases and solutions, and delivering better experiences, enabling more people to truly utilize the purchased models or open-source models.

Wall Street News has learned that Zhipu has partnered with collaborators in finance, education, healthcare, government affairs, and enterprise services to jointly advance the application of Agentic LLM. In February of this year, Zhipu officially announced a collaboration with Samsung based on Agentic LLM, bringing the Agent experience to Samsung's latest Galaxy S25 series smartphones. At the same time, Zhipu has also successively reached cooperation agreements with cities such as Beijing, Hangzhou, Shanghai, Chengdu, and Zhuhai.

Zhipu is also actively expanding overseas. On the same day, led by Zhipu, a "Self-Sufficient Large Model International Co-Building Alliance" was officially established, initiated by ten countries from ASEAN and along the Belt and Road, to help Belt and Road countries establish autonomous AI and build controllable national-level AI infrastructure.

Regarding commercialization, Zhipu achieved an overall growth rate of over 100% last year, with many leading industries already making inroads, resulting in a certain scale effect. For this year's expectations, Zhang Peng stated that after another round of popularization, the market will see more than tenfold growth, presenting greater opportunities.

"The entire model and commercial path will undergo some changes, and we will make adjustments. However, we will still maintain a consistent and stable pace and effectiveness in commercialization, continuously improving the results of commercialization," Zhang Peng stated.

When discussing the company's current strategic focus, Zhang Peng told Wall Street News that Zhipu positions itself as a technology-driven company, with another leg being the commercialization path. These two legs are not contradictory or competing for resources; they are in a dynamic adjustment process.

Zhang Peng further stated: "The advancement and evolution of technology have reached a point where it must be deeply integrated into industries and applications to absorb nutrients and feed back into technological research and development. Therefore, since last year, we have been vigorously promoting industrialization and commercialization, which will require some resource investment. However, from the overall perspective of core tasks and resource investment, we are still investing more resources in technological research and innovation."

This race towards AGI is still in its early stages. For Zhipu, although it has already explored the L3-Agentic LLM stage, the road ahead is still fraught with challenges. To achieve greater innovation on a global scale, it must go all out