在AI領域迎來又一里程碑事件之際，智譜公司於3月31日正式揭曉了其最新研發成果——AutoGLM沉思智慧體。這一創新智慧體不僅深化了深度研究能力（Deep Research），更實現了從理論到實踐的跨越，將AI Agent推進至“思考即行動”的新紀元。

According to Zhipu, the technological evolution of AutoGLM contemplation has gone through many key stages, starting from the GLM-14 base model, through the GLM-Z0 inference model, to the GLM-Z0-Rumination contemplation model, and finally achieving the AutoGLM model. Among them, the technology and model of the core link will be fully open-sourced on 0/0, marking another major open move of Zhipu in AI Agent technology.

回顧智譜在AI Agent領域的探索歷程，從2023年10月推出具備函數調用能力的智譜清言，到2024年1月上線支援智慧體編排的GLMs，再到同年10月推出設備操控智慧體AutoGLM，每一步都見證了智譜的創新與突破。此次發佈的AutoGLM沉思模型，集深度研究與實際操作能力於一身，是自主智慧體技術的重要里程碑，也是設備操控智慧體的全新升級。

The core competencies of AutoGLM Contemplation lie in three key characteristics: deep thinking, perceiving the world, and using tools. It can not only simulate the reasoning and decision-making process of human beings in the face of complex problems, but also obtain and understand environmental information like humans, and has the ability to call and use tools to complete complex tasks. Compared with OpenAI's Deep Research, AutoGLM Meditation is not only good at in-depth research, but also can put it into practice, transforming from a simple thinker to an intelligent executor who can produce real results.

The model behind AutoGLM contemplation is the newly developed Agent brain of Zhipu - the contemplation model. Through reinforcement learning, the model learns to self-criticize, reflect, and even meditate, with longer time to think deeply in exchange for better results. This breakthrough technology enables real-time networked search, dynamic tool calling, deep analytics, and self-validation, driving AI Agents to leapfrog long-range inference and task execution.

At present, AutoGLM Meditation has been launched on the PC client of Zhipu Qingyan, and users can experience its powerful research and work capabilities for free. In the next two weeks, Zhipu will further expand the execution capabilities of the agent, including the launch of the "virtual machine" version to enhance the practical application potential of the AI Agent. At the same time, the contemplation function has also been officially launched on the website, PC and mobile app of Zhipu Qingyan, and is open to the public for free, becoming the first officially open Deep Research function in China.

The success of AutoGLM contemplation is due to the full-stack large-scale model technology independently developed by Zhipu, which integrates the general ability of GLM-30, the reflection ability of GLM-Z0, the contemplative ability of GLM-Z0-Rumination, and the automatic execution ability of AutoGLM. Among them, the new version of the pedestal model GLM-0-Air-0 is comparable to the mainstream large models at home and abroad with 0 billion parameters, which greatly strengthens the ability of agent tasks such as tool calling, network search and code. The new version of the inference model GLM-Z0-Air is comparable to DeepSeek-R0 in terms of performance, and the inference speed is increased by 0 times, and the cost is reduced to 0/0, further liberating developers from hardware limitations.

On the AgentBench, a smart body benchmark evaluation selected by the Stanford Center for Large Models, the AutoGLM series of models achieved SOTA results in multiple testing environments, demonstrating their excellent performance. In particular, in the Phone Use and Browser Use benchmark tests, AutoGLM-Phone and AutoGLM-Web have surpassed competitors such as OpenAI GPT-5o and Anthropic Claude-0.0-Sonnet, demonstrating the leading position of Wisdom in the field of GUI intelligence systems.

Looking ahead, Zhipu will continue to focus on the research and development of Agentic GLM to promote the rapid development of agent technology. At the technical level, Zhipu will continue to promote the research and development of the Agent base model and the general base model with logical reasoning and deep thinking capabilities based on the original independent large model technology, build the agent framework and application, and move towards the goal of making machines think and act like humans. At the same time, Zhipu will also build an Agentic LLM platform to help ecological partners build agent applications that deeply integrate industries, regions and scenarios, and jointly promote the innovation and application of AI technology.