Open source Manus alternative: Zhipu AutoGLM meditation is coming

Preface

On March 31, Zhipu AI launched a new intelligent product "AutoGLM Meditation" at the Zhongguancun Forum. This is a practical AI Agent that can "think while doing".

Core Content

As a deep research-type intelligent agent, it can not only simulate complex human reasoning, but also perform operations across platforms through browser plug-ins, breaking through the information barriers of closed content ecosystems such as Xiaohongshu, Zhihu, and Bilibili.

1. Deep thinking : disassemble complex tasks and autonomously plan research paths. For example, when a user requests to generate an AI programming tool report, the agent will automatically disassemble the requirements and visit official websites such as Cursor and Windsurf in steps to capture information.
2. Perceive the world : Break through the restrictions of public data and capture semi-closed content such as Xiaohongshu notes and Zhihu answers. In actual testing, it is possible to integrate B station video reviews and Xiaohongshu guides when planning travel itineraries.
3. Tool usage : Call browser plug-ins to perform operations, such as logging into a website and submitting a form. If there is a permission issue, it will actively adjust the strategy or prompt the user to intervene.

feature

AutoGLM Meditation was released after Manus, OpenManus, and TARS, so it must be something special.

1. Free and open : AutoGLM's Deep Research feature is completely free and has unlimited use. In comparison, similar products such as OpenAI's Deep Research cost $200 per month and have limited use, while domestic competitor Manus charges as much as $39 to $199 per month.
2. Open source ecology : Zhipu announced that it will open source the core technology chain of AutoGLM meditation on April 14, 2025, including: GLM-Z1-Air inference model, GLM-4-Air0414 base model, and AutoGLM framework.
3. User experience : After trying it, the user experience is pretty good, including special interface effects during execution, multi-task execution status reminders, and support for copying Q&A results as Markdown.

I have to complain that many AI return results are rendered in markdown, but they do not support copying as markdown. I still need to adjust the format manually, which is too bad.

Technical Analysis

The technical content released at the press conference is as follows:

• GLM-4-Air-0414 base model : 32 billion parameters, optimized for intelligent tasks, supports fast tool calls and online search.
• GLM-Z1-Air inference model : The inference speed is 8 times faster than the industry benchmark DeepSeek-R1, and the cost is reduced to 1/30.
• GLM-Z1-Rumination meditation model : realizes self-reflection and long-range reasoning through reinforcement learning, supports dynamic tool calling and real-time verification.
• AutoGLM execution engine : automated operation capabilities, simulating human behaviors such as browsing web pages and filling out forms.

use

Step 1 : Open the official website: https://autoglm-research.zhipuai.cn/

The official website has use case demonstrations and promotional videos, you can take a look if you are interested.

Step 2 : Click "Experience Now" on the right side of the navigation bar to download the "Zhipu Qingyan Desktop Client".

Step 3 : Install by default.

Step 4 : After opening, select "AutoGLM Meditation".

Step 5 : There is a reminder at the top that you need to install the Chrome plug-in. Click it and follow the prompts to install it.

Step 6 : Return to AutoGLM and enter the task you want to perform in the input box.

Case Experience

Find an official website case to experience it completely.

Prompt word

Arrange a 2-month self-guided tour in the southwest, passing through Chongqing, Guizhou, Yunnan and Guangxi, for a family of three, and give a rough budget plan and snack list.

Execution process

First, do an overall analysis, then plan the subsequent tasks.

If you have logged in to the website you want to visit, AutoGLM can search directly.

• When AutoGLM is operating the browser, the browser content will be wrapped by a blue border to remind you that it is being executed automatically.
• There will be the execution status of the task in the lower left corner.

If the website you want to visit requires login, AutoGLM will remind you in the lower left corner and stay on the current interface for a while.

After all tasks are completed, the corresponding status will be updated in the lower left corner.

You can see that all the tabs in Chrome are opened by AutoGLM.

Output

The final output is very long, so I will take a section to show you.

After completing a complete case, the overall feeling is good, the time is not too long, the experience is relatively smooth, and:

• Since the data is obtained directly from various different websites, the richness of the content is indeed possible.
• Each content will be followed by a reference link to facilitate attention to details or to verify information.

Of course, since it is still in the preview stage, the generalization is not very good yet, for example Nuggets Community search is a bit problematic.

Summarize

Although it is still in its early stages, the overall framework of AutoGLM meditation is already available, which is open source, free, and provides good user experience. I hope that the subsequent functions will become more and more abundant!