Qwen's move is more meaningful than releasing a large model

A few days ago, Ali Qwen chat launched "Deep Research", claiming to be able to compress complex tasks that take several hours to be completed in tens of minutes, and it is open to all users for free.

For someone like me who often needs to collect field information, it is very attractive, so I experienced it for two days as soon as possible.

Open the QwenChat interface and you can find the entrance to in-depth research in the function options; the official said that an intelligent assistant system integrates a large number of online information sources and can plan complex tasks.

When I used it for the first time, I entered a relatively broad research topic: "Please help me analyze the current competitive landscape and future trends of AI search products." After submitting, it did not start searching immediately, but confirmed my needs first.

The whole process is divided into three steps:

First, the system immediately gave three major questions, which included several minor ones. It seemed that it was planning the report structure and knew how to break down a major requirement into small tasks.

In the second step, Deep Research starts to automatically search, filter and integrate relevant information on the Internet. The whole process is automated and the internal logic is understood.

In the final step, it makes further adjustments based on the results of the phased output, making it more like an intelligent entity with both "learning" and "adaptation" capabilities.

In about 10 minutes, a research report on "Current Competitive Landscape and Future Trends of AI Search Products" was generated; the overall report has a clear structure, point-by-point explanation, comprehensive content, and comes with reference links; although it is generated by AI, it is quite readable.

However, there are advantages and disadvantages. What are the disadvantages?

Like other in-depth research products, everyone now likes to see who has longer research reports; it’s like giving it a topic and it keeps writing, which feels like challenging the system to maximize its input capacity.

Although the report is long, the quality is not up to par because there are too many words but not much truly useful content.

For example, I read a report of 10,000 words and found that the most noteworthy part was only about 3,000 words. The rest was either pure theory or boring written language that was difficult to read.

There is also a third problem, it gives people an illusion of "seemingly correct but inaccurate". What it writes seems to make sense, but I don't know if it is accurate; it can put a lot of text together and read smoothly, but as I read, I feel something is wrong.

To mention one more point, macro-narrative will write a lot of content in a broad sense, but it cannot go into specific details. It is like telling a big story, but it only simply outlines the outline but does not explain the plot clearly.

These past two days, I have been thinking, who is Qwen's Deep Research feature suitable for? It is probably suitable for three types of people:

The first is that ordinary people make life decisions.

You want to enroll your child in a summer camp, but there is too much information online and it’s hard to choose. Deep Research can help you find information about all institutions at once: which one has a good reputation, what the price is, and what the course content is, so you don’t have to look through them one by one.

Then, there are people who write, especially those who are interested in literature and philosophy.

This type of content is not like a research report, which does not require so much data support; Deep Research is very useful. You can let it help you look up information, find inspiration, and organize your thoughts, making it a lot easier to write.

The third category is students and researchers. If you want to do a project, it takes a lot of time to search for literature, make an outline, and write a review. You can use it to quickly collect information, sort it, and help you build a framework, which will speed up the whole process.

So, how does it do this? What is its underlying logic?

It could be something like this:

First search for information online, pick and sort it, and keep the truly useful information. This step is similar to searching for things online, but it is faster and more accurate.

Then, it will call on many external knowledge bases , such as Wikipedia, academic papers, and even some internal data systems of enterprises. These knowledge bases are like the brains behind it.

Moreover, it does not get everything done at once, but will constantly adjust during the process. If the information in a certain direction is not complete, it will check it again, or change the way of thinking to continue the analysis. The technology of reinforcement learning should be used behind this process.

Finally, it organizes the complex analysis results into a language that ordinary people can understand and generates a well-organized report.

If I were to summarize it in one sentence, Qwen's Deep Research connects search, reasoning, execution, feedback and expression together, like a general agent.

After two days of in-depth experience, I feel that it cannot completely replace Manus at present.

Every time Manus creates a new task, if there is a new inspiration during the execution, just interrupt it and input it. It will immediately adjust its direction and re-search and retrieve relevant information. This ability to "make changes while doing" is quite similar to the feeling of working with people.

Another feeling is that when I used Manus before, although its output was very comprehensive, in many cases the amount of information was too large and the key points were not highlighted.

Recently, it seems that the length control has been optimized and the accuracy of the content has been improved. In other words, it no longer pursues "writing more" blindly, but pays more attention to "writing more accurately".

I made a temporary interjection while it was performing its task yesterday:

What is the difference between Quark Deep Research and Grok 3's Deep Research? It immediately started to look up information and sorted out the core differences between the two. The entire text is not long, the point of view is clear, the structure is clear, the information is also in place, and there is little redundancy.

Judging from these details, if large domestic manufacturers can work harder in the direction of "flexible adjustment during tasks", the final generated report may be closer to real needs and have more practical value.

This also exposes a problem: everyone has different understandings of deep research. The core difference lies in: is it a "task assistant type" or an "information retrieval type"?

What is a Task Helper?

Manus is more like a researcher who can think. No matter how I ask you to disassemble it, how to do it, how to fine-tune it, how to change the direction, or how to add questions, it can handle them all.

Qwen's current Deep Research also has a certain ability to understand and adjust tasks, but overall it is more like a tool that "can plan + organize information."

It can help you look up information, make outlines, and organize your thoughts, but in terms of the sophistication of task decomposition and the depth of interaction, it does not reach the "human-like research process."

Therefore, some products say that they have Deep Research functions, but in fact they can search deeper and integrate better; while the Deep Research of some products can really "do research" with you, making changes as you go, and becoming more and more accurate.

These two differences are critical and determine whether you can really use it to replace part of the manual research work.

I also discovered a trend: the Deep Research functions of many model products for C-end users in China are basically moving in the direction of Grok3.

What does it mean?

They pursue "complete information and long output", hoping to cover all aspects related to a problem and restore a complete information picture as much as possible; the core of this idea is "comprehensiveness", which does not mean that you can only see one corner, but try to see the whole puzzle.

I think this is a good idea.

A few days ago, I was going to write an article about Pang Donglai and wanted to check two data: What is the average salary of employees? How much does the management get?

I used some AI search tools, and the data provided by each one was a little different, which was very confusing; but when I used Quark's Deep Research to check, it not only gave the answer, but also listed the data from different sources and different time periods, and even marked the source of the information and the calculation method.

The most important thing is that it helps me filter out a lot of obviously unreliable information, organizes the content with a high probability of being accurate together, and then lists the controversial, uncertain, and low-probability information separately, like giving a "bird's-eye view" that allows me to see the ins and outs of the entire problem at a glance.

So, I think that Deep Research is now more like a deep retrieval system that helps you filter, integrate, and present the most valuable parts while not neglecting the details that deserve attention.

As for actually becoming a task assistant, that is another category.

Therefore, it is important to distinguish between the two. Platforms like ByteDance's "Button Space", Baidu's "Xin Xiang", 360's "MCP Universal Toolbox", Alibaba's Yunbailian, etc. are more like task planning and execution platforms.

If you need an AI that can help you "do things", you should pay attention to this type of task assistant platform; if you just want to quickly understand the ins and outs of a problem, then the Deep Research function is enough.

Seeing this, some people may ask another question: What is the difference between Deep Research and Deep Thinking?

This concept confuses many people.

Deep Research, literally means "deep retrieval".

You lost your key, so you start searching the whole house, looking here and there, even shining a flashlight in every corner, and finally you find it. Although this process is tedious, the goal is clear: to find the key thing.

This is actually the core of Deep Research: the ability to dig deep.

What about Deep Thinking ? It’s a little different. After you find your key, you suddenly think of a question: Why do I keep losing my keys? At this point, you have entered the stage of “Deep Thinking”.

You might start to wonder: Am I in too much of a hurry to get out? Should I get a bag with a reminder function? Should I change this bad habit? You might even think of other similar questions: "Do I lose everything easily?"

This is deep thinking: it is not just about looking for something, but about analyzing the causes, reflecting on the behavior, and proposing improvement plans before and after finding it.

Simply put, Deep Research is the process of finding the key; deep thinking is when you find the key and start thinking about why you lost it and how to avoid it in the future. One is "finding" and the other is "thinking it through."

As early as February 2025, after OpenAI launched Deep Research, major foreign companies quickly followed suit.

On February 14, Perplexity also launched its own Deep Research function; almost at the same time, Grok-3 also integrated DeepSearch capabilities.

There was almost no movement in the country at that time.

It was not until the first half of this year that ByteDance, Baidu, Tencent, and Alibaba launched deep thinking related functions, which seemed to be a catch-up. But in fact, this wave of Deep Research capabilities has not yet fully caught up, or even been ignored.

I think that in the field of AI research, the gap between China and abroad has not yet been completely narrowed, and the big companies really need to work harder.

Because the amount of information is getting bigger and bigger, it is easier to confuse people. A lot of content is generated by AI, and even when AI itself quotes, it is unclear: is it citing real information, or is it copying another AI-generated content?

I recently used Tencent Yuanbao and clicked on a lot of content, but I couldn’t find the original source at all. If you don’t check carefully, you might mistakenly think that this information is reliable.

True deep search, however, aims to solve this problem by finding truly reliable and valuable content through deep digging capabilities.

Qwen has taken this a step further.

This step may seem small and imperfect, but it is of great value. It is more practical than simply releasing a model because it solves problems that users will actually encounter.

So, the question is: will it extend this capability to other products? What impact will this step have on its ecosystem? I am still observing further.