Zhipu launches AutoGLM, OpenAI no longer has exclusive access to Deep Research

Quick, this moment comes too fast,

I was just attending a press conference in Zhongguancun!

Deeper Research is no longer exclusive to OpenAI

Zhipu has fully adopted Deep Research, which has AutoGLM capabilities.

AutoGLM Meditations

Combining the deep research capabilities of the inference model with the browser usage capabilities of the Agent,AutoGLM MeditationsIt can easily access the web pages that need to be logged in in the existing AI search engines (hownet, certain books, public accounts, certain JD, etc.),

It can also output structured research reports. The continuous operation time of a single task easily exceeds 15 minutes, and it can solve complex tasks related to information search in Manus. The key is that the free trial is unlimited.

Unlimited, unlimited, unlimited, important things to say three times!

OpenAI's Deep Research is done 10 times a month, Gemini2.0 Deep Search is done 10 times a month, and Manus complex tasks are done 6-7 times a month.

In fact, we have already developed Auto-GLM for mobile phones, GLM-PC for computers, and Qingyan for browsers earlier. It is not surprising that Zhipu was the first to do it.

During the 72 hours of internal testing of Zhipu's new Agent, it took over my phone, computer and browser

We will put the usage and principle behind.

Now let me experience it in depth.AutoGLM MeditationsI gave it a hell level of difficulty, with a lot of food to satisfy your hunger. Here we go!

PS: The case tips and sharing links below are all packaged, just send me "Meditation" on WeChat

Q1: I need a 7-day Japan trip from Beijing from April 15 to 23, with a budget of $2500-5000, for my fiancee and I. We love historical sites, hidden treasures, and Japanese culture (kendo, tea ceremony, Zen meditation). We want to see the deer in Nara and explore the city on foot. I plan to propose during this trip and need a recommendation for a special location. Please provide a detailed itinerary and a simple HTML travel guide with maps, attraction descriptions, essential Japanese phrases, and travel tips for us to refer to during our trip.

This question is complex. Due to time, budget, preferences, and uncertainties, the model needs to choose a suitable place to propose from the entire itinerary. There are also additional requirements, such as maps, attractions, and language phrases.

The whole journey takes 18 minutes.AutoGLM MeditationsIt perfectly fulfilled my needs. The video is 2x faster. I strongly recommend you to take a look at the details of the interaction:

1. Supports mutual verification of different information sources (such as certain websites and certain books), which is very suitable for my daily travel strategy.
2. You can actively select when previewing informationWithin one month,Hottest,Read only articlesand other conditions
3. The operation speed of a single page is very fast. When encountering a page that does not meet the search conditions, it is almost skipped in seconds, without entanglement or jamming.
4. The additional clicks, slides, and other operations in the middle will not affect the subsequent automatic operations. If you are not in front of the computer, you don’t need to worry about being stuck on the login interface. It will reserve a suitable time. If it doesn’t work, just change the information source.

My overall feeling isComfortWhen operating the web page, the lower right corner will actively feedback what the model is doing, so you won't feel like you're waiting in vain. I only switched pages a few times to return to the dialogue interface.

The preview records will be kept on the computer throughout the process, and I can also bookmark the website or do additional browsing.

The second example is more difficult. It requires not only a broad search, but also a deep search to understand the scope I outlined.

Q2: Compile a detailed list of key global AI influencers, including opinions from leading AI company executives, investors, and researchers on Deepseek R1.

This question also has traps. First, I want a global scope, and the English website must be operational. Second, I want a specific thing.R1, so the requirements will naturally be higher.

It took 18 minutes.AutoGLM MeditationsI was given a serious 5,000-word report. Unlike the first case, it can do secondary reasoning. Specifically, it uses the first round of information that does not pass through the web page operation as the source, and jumps to multiple web pages of the same website like the growth of tree roots. Of course, it also cleverly jumps to the DeepSeek website to read the logs sent out. . .

Let’s take a look at the quality. The structure of the research report starts with the analysis of comments, including academicians, MIT Technology Review, AI experts Yann LeCun, Geoffrey Hinton, Meta, Microsoft, NVIDIA, etc., covering the evaluations of academia, investment circles, and even highly praised posts in a certain book. This search method can only be done by Deep Research, and a comprehensive evaluation of an object can be obtained in one sentence. Before this, I had to look for paper reviews or spend three or four days looking at wikis.

Q3: What vertical search AI solutions exist in the fashion industry? In what specific scenarios are they used? What are their pricing models? Which parts of the value chain do they serve? How do these products differentiate from each other?

In the unpopular type of questions,AutoGLM MeditationsCan you still maintain your high standards?Industry SolutionsThis type of question usually only appears in a certain white paper. Here I intercepted part of the output, without changing a word:

• AI applications in the fashion industry include: fashion design, recommendation systems, conversational robots, trend analysis, intelligent manufacturing, etc.
• Several specific companies and products were mentioned, such as Heuritech, Bosideng AI system, Fashion Diffusion big model, Style3D AI industrial model, etc.
• Heuritech is a company that uses AI technology to analyze social images to predict retail trends, analyzing more than 3 million social images every day.
• The application of AI in the fashion industry includes virtual try-on, fashion synthesis, fashion retrieval and other functions

The benefit of the inference model isAutoGLM MeditationsIn each section of a complete conversation, I review what I have done so far, what needs to be done based on the current status, and the differences between different information sources. The coolest thing is that I even reveal the pricing plan.

AutoGLM MeditationsThe appearance of also solved a big problem for me.

There must be everyone who has not heard this question, " What is the best XXX? " This XXX is very creative. It can be a large model, AI raw picture, AI video, AI PPT and so on. However, everyone is different and everyone has different backgrounds. It is difficult to get a satisfactory answer through search.

Q4: As a self-media creator with more than 3 years of experience focusing on video production (shooting/editing) and narrative design, I need you to get some tools to simplify script writing and plot development.

Main requirements: - Dialogue/storyboard script creation tool

Thanks to the inference model,AutoGLM MeditationsIt can keep my background information in every round, and it sets a plan for itself right from the start:

1. Gain insights into professional screenwriting tools (both traditional software and AI tools)
2. Research storyboarding tools and software
3. Explore narrative structure tools designed for video creators
4. Find a comprehensive solution that simplifies your script-to-video production process
5. Learn about the actual usage experiences and recommendations of professional creators (especially in the fields of video and narrative design)

The biggest difference from searching by yourself is thatAutoGLM MeditationsThe results take into account both AI tools and traditional tools. I never expected that DeepSeek, Midjourney, Celtx, Final Draft, and StoryBoard Quick would be put together one day.

The final result takes into account the applicable population, price range, scenario recommendations, and whether the tool has built-in AI assistance.

The last case is a panacea.

The key to learning something is to be systematic, especially learning a language.AutoGLM MeditationsThe problem with full performance is:

Q5: Can you create resources for me to learn language telugu?

This Telugu language is spoken primarily in southeastern India, with the first written material in the language dating back to 575 AD.

AutoGLM MeditationsFound it in the first round50languages,Tsinghua University Language Teaching Center,Shanghai International Studies UniversityAnd other resources, much smarter than me always thinking about a certain disk.

It also provides a pretty good language learning system, including online courses, vocabulary lists, test systems, pronunciation guidance, and even my own secret tips for learning languages: finding an affordable oral teacher. I also deliberately sent it to a few of my friends who do online education, and their responses were quite consistent:

This is a pretty good start!

In my opinion,AutoGLM MeditationsThis means that there is no need to design complex prompts.

It can directly display the web pages it browses and save them on my computer for me to read more or save them.

I can also directly use my personal account to search and learn the content recommended by the platform that meets my personal preferences.

There is no need to worry about the model missing out on unpopular or non-intuitive information.

Overall, it is more comprehensive and understands me better.

At the press conference, the Zhipu team also announcedAutoGLM MeditationsI created a knowledge sharing account for a certain book, and it already has over 3k followers... (I feel a sense of crisis)

 How to use?

The usage of the full-blood version is quite friendly.

The first step is to install the PC version of Zhipu Qingyan and install the Qingyan browser plug-in

?  autoglm-research.zhipuai.Cn/

The second step is to selectAutoGLMThen ask directly!

The method to verify whether the installation is successful is very simple, because Zhipu has provided many cases. You can select any one and send it to the conversation.AutoGLM MeditationsThe web page operation will pop up automatically.

Unlike some Manus-like projects that rely entirely on the capabilities of Claude and GPT,

AutoGLM MeditationsThe large models supported by the backend are all self-developed

Continuing the consistent style of Zhipu, the inference model usesGLM-Z1-Air, providing Agent with reasoning, planning and reflection capabilities. The effect is comparable to DeepSeek R1, the speed is increased by 8 times, the price is 1/30 of R1, and it can run on consumer-grade graphics cards;

The base model isGLM-Z1-Air-0414, has been specially trained for reasoning and agent tasks, and is more proficient in agent tasks such as code writing and tool calling

Based on  GLM-Z1-Air, the Rumination model GLM-Z1-Rumination was trained through reinforcement learning , which improved the model's ability to complete long-range reasoning in combination with tools.

Therefore, the evolution line of Zhipu’s model is clear and very targeted.

 Last words 

Our source divine wisdom spectrum has never let you down.

The previously announced year of open source is still in full swing.

Only 27 days have passed.AutoGLM MeditationsBase modelGLM-4-Air0414,Inference Model GLM-Z1-Air,Rumination Model Z1-rumination, intelligent agents and other core frameworks will be officially open sourced on April 14.

There will be a free version of the base model that will be upgraded toGLM-4-Flash-0414, my list of commonly used models can be upgraded again.

Deep Research has been launched for 56 days. It is very powerful. Deep Search and Deeper Search that followed it are also very powerful.

But how many people have used it?

It's because their threshold is too high and the number of times you can experience it is too small.

But Zhipu has been constantly lowering the threshold.

Isn’t this technological equality?

Doesn't this give everyone the opportunity to use Agent?

I always think that

This is a future development.

Especially with the development of AI era,

The most essential and contributing step.