Wanlian Securities: Zhipu releases AI Agent AutoGLM for reflection, entering the year of "open source" for Zhipu

Wanlian Securities released a research report stating that the AI Agent product AutoGLM, newly released by Zhp, is expected to drive technological progress in AI Agent products and further upgrade applications on the edge. Compared to other Agent products, AutoGLM has three major features: 1) the model is based on Zhp's Z1 reasoning model reinforcement learning training; 2) the model has GUI reading capabilities; 3) the model integrates Zhp's proprietary AutoGLM device operation capabilities. At the same time, Zhp plans to officially open source its core chain model and technology, which will help accelerate the overall ecological construction of the AI Agent industry and the application of AI. Wanlian Securities' main points are as follows: Event: On March 31, 2025, Zhp officially released AutoGLM Rumination at the Beijing Centergate Technologies Forum. According to reports, this is the first Agent that combines deep research and practical operational capabilities. Currently, AutoGLM Rumination is online on the Zhp Dialogue PC client, and users can experience its research and operational capabilities for free. The version released this time is a preview version, core support for research scenarios; rumination functions, which have been officially launched on the Zhp Dialogue webpage, PC side, and mobile app, are free and unlimited for everyone. This is also the first officially open Deep Research feature in China. AutoGLM Rumination drives the AI Agent into the "think and act" stage, with three key characteristics AutoGLM Rumination is an autonomous intelligent entity (AI Agent) that can explore open problems and perform operations based on results. It not only has deep research capabilities, but also practical operational capabilities, achieving the goal of "thinking and acting at the same time." AutoGLM Rumination integrates three major capabilities: 1) Deep thinking: simulating human reasoning and decision-making processes when faced with complex problems; 2) Perceiving the world: being able to acquire and understand environmental information like humans; 3) Tool usage: being able to call and operate tools like humans to complete complex tasks. Compared to other Agent products, AutoGLM Rumination has three main features: 1) the model is based on Zhp's Z1 reasoning model reinforced learning training, similar to Open AI's Deep Research training process based on the o3 model, the model can autonomously plan and make dynamic decisions based on different task goals, continuously adjust plans based on feedback without the need for pre-designed workflows; 2) the model has GUI reading capabilities, not only relying on API calls, but also being able to open and browse web pages like humans, browsing dozens of web pages for up to tens of minutes to find the information you need; 3) the model integrates Zhp's proprietary AutoGLM device operation capabilities, such as being able to send emails based on the results after completing a report. Zhp continues to lead innovation exploration in the AI Agent industry and is about to usher in the "open source year" of Zhp. In the past research and exploration of AI Agents, Zhp has always been at the forefront, continuously releasing innovative products. From the previous releases of Zhp Dialogue and GLMs, to the world's first device control AI Agent AutoGLM launched in October 2024, to the three products based on AutoGLM upgraded for the web, mobile (Android), and PC sides at the Agent OpenDay at the end of November 2024: AutoGLM Web, AutoGLM (Android), and GLM-PC, Zhp has been leading the innovation exploration of AI Agents. The release of AutoGLM Rumination reflects Zhp' core understanding of AI Agents as "not only able to think, but also able to take active actions" and promotes AI Agents from mere thinkers to intelligent performers who can deliver results. The technical evolution path of AutoGLM Rumination includes: GLM-4 base model GLM-Z1 reasoning model GLM-Z1-Rumination rumination model AutoGLM model. Among them, the GLM-Z1-Rumination model breaks through the limitations of traditional AI relying solely on internal knowledge reasoning and innovatively combines real-time online search, dynamic tool calls, deep analysis, and self-validation to form a complete autonomous research process; the AutoGLM series models have also made important progress, being selected in the Stanford Model Center's "AI Index 2024" intelligent agent benchmark evaluation AgentBench, achieving SOTA results in five testing environments. Zhp plans to officially open source the core chain model and technology on April 14 to promote industry ecological development. Risk factors US-China technology friction; domestic AI large model competitiveness lower than expected; AI Agent application landing lower than expected; AI industry technological development and computational power demand lower than expected.