This article is crafted based on publicly available information and is intended solely for information exchange purposes, not constituting any form of investment advice.

Manus ascended to the pinnacle of AI trends seemingly overnight, propelled by a relatively informal product launch. On certain second-hand platforms, invitation codes for Manus have skyrocketed to 50,000 yuan. Unsurprisingly, two groups are now experiencing heightened anxiety: investors and large corporations.

What is the allure behind Manus? Is it merely a flash in the pan, or does it genuinely hold substantial value?

01 What Can It Do?

Manus can autonomously complete a comprehensive financial report analysis for Tesla, generate a fully interactive website, or even craft an RPG game ready for immediate play. It can set up a development environment, write code, debug, and compile it all by itself. For those interested in creating a regional population research report, Manus can access databases and search autonomously to determine the most relevant data for the task.

Remarkably, Manus can execute a complex analysis project with a single prompt. According to its official website, its performance surpasses OpenAI's DeepSearch under the GAIA benchmark, the evaluation standard for AI agents.

After analyzing user cases shared on Twitter and domestic social media, I found that Manus is on par with DeepSearch in terms of execution process. Each step's execution code and output files are fully visible and neatly categorized for user review. When it comes to financial report analysis and valuation results for Tesla and NVIDIA, Manus holds its own against seasoned investors on Xueqiu.

What truly impresses me is Manus's ability to execute a series of complex tasks, such as data acquisition, computation, development environment setup, and testing, all through its own virtual computer. This virtually eliminates the need for users to engage in complex foundational work, enabling even newcomers to start from scratch with minimal effort.

02 The Team Behind It

Manus's buzz is not solely due to its high-quality output but also because it is the work of a locally grown domestic team, igniting excitement on social media. In the AI field, domestic teams are keeping pace with the US dream team that boasts global talent.

Manus is another product from Monica AI, a domestic AI startup. Unlike nascent foreign chatbots, Monica AI doesn't just provide an entry point for using a model; it offers a plethora of directly usable API interfaces for various verticals. Users can start using it immediately without worrying about prompt tuning.

Monica AI's founder, Xiao Hong, recognized the limitations of simple chatbot products in facilitating interaction between large models and users. While the model is excellent and possesses certain 'superpowers', capable of completing complex projects with adjustments, users needing to complete tasks can only feed them to the model in batches through 1v1 conversations, resulting in step-by-step outputs. Worse still, users may not be able to confirm if the tokens input to the model are the optimal solution for the task.

Manus allows users to continue using the dialog box to solve more complex, high-intelligence tasks in one step. From this perspective, Manus resembles a collection of multiple planned AI tasks.

The team deliberately shuns showcasing the magical aspects of large models and instead stands from the user's perspective, contemplating how to leverage the model's advantages to complete daily tasks. This is a pragmatic and valuable exploration.

Manus's story underscores serial entrepreneurs' precise grasp of user needs and the team's exceptional execution ability. We will delve into the technical prowess required for AI later.

03 Product Highlights

According to articles from "Cyber Zen Mind," a self-media close to the Manus team, the cost of a single task is around $2, which is one-tenth of Deep Search's cost. Built on Claude 3.7, Sonet, and the domestic Qwen model, Manus has added its own training process. The founder revealed in a podcast interview that even when working on Monica, they didn't merely serve as a prompt transit station, directly throwing user needs to the model, but instead added a product tuning process.

Just like a martial arts prodigy with exceptional talent and unique bones, they need to work hard to become the best in their field.

Manus's ability to independently plan, think, and complete tasks stems from its hybrid model architecture. Based on Claude and Qwen, it doesn't mean that one aspect of the task is handled by Claude while another by Qwen. Manus breaks the 'boundary' of the models, allowing each to excel in its respective task aspects.

Combining multiple large models with Manus's virtual cloud execution environment enables users to complete complex needs that would typically require multiple rounds of chatbot conversations with just one keyboard stroke.

A Guess About the Manus Execution Process:

The term "engineering-driven innovation" mentioned in the "Cyber Zen Mind" article suggests that Manus employs a significant amount of engineering methods combined with model capabilities to complete tasks during its execution process. Analyzing multiple social media case sharings, we made a simple guess about Manus's operating mechanism.

After analyzing the input through a large model (this step guesses it's Qwen), Manus first determines the task type. Internally, Manus should have a categorization of inputs, such as programming tasks, multimodal content generation, task planning, advisory categories, etc. Secondly, if the input language is English, it's guessed that the task execution may primarily be Claude-based.

Specifically, at the task execution level, the guessed process is as follows:

1. Upon receiving an input, the large model first classifies the task to determine the model selection for subsequent execution.

2. The large model then breaks down the input into multiple sub-tasks with hierarchical relationships, priorities, and execution sequences.

3. After sequential execution, the output of the superior task becomes the input for the next task.

4. The outputs of multiple tasks are combined into a final analysis result.

For the selection of which model to execute each sub-node, we guess that if it's a programming task, Claude will primarily handle the execution, while if it's a Chinese decision-making or advisory task, Qwen will take the lead. Depending on the complexity of the sub-node task, a single task during execution may even be a mix of Qwen and Claude, followed by a model comparison of the results, ultimately retaining the one with the best effect.

From the above analysis, it's evident that the initial requirement's input quality determines the subsequent task type judgment, as well as the sub-node planning and execution efficiency. Therefore, Manus has high requirements for input quality, typically requiring users to describe the requirements in detail and completely (the more detailed, the better).

This may become a point of criticism for Manus. After all, the execution of an input is lengthy, and if the description isn't clear and detailed enough, it will lead to a waste of time and computing power, seriously impacting the product experience. This reflects the product's immaturity, and I believe the team will improve this in the future, considering that a single task execution costs as much as a cup of bubble tea.

04 Current Issues

From 2022 to 2025, people have experienced over three years of cognitive shocks from the large model trend. Users' experience with large model products has generally passed the "AHA Moment" stage that determines a product's survival. A large model product must not only be amazing in effect but also stable and fast to retain users in the long run.

Effect Stability

Looking back at Manus, given that our Manus account is still on the wishlist, feedback from users on Twitter and domestic social media slightly differs from the widespread praise for DeepSeek. Users are not optimistic about the stability of Manus's output, and there are even some hilarious moments where facts are misinterpreted.

If the data or facts for a task are used incorrectly during execution, it will render the final output untrustworthy, thereby wasting the $2 cost.

Computing Power

Besides inconsistent performance, another criticism is the excessively long execution time for a single task. Although Manus has demonstrated its task execution logic and process, users still have to endure a long wait. One user tweeted that their task took 4-5 hours to execute.

Behind this is the team's underestimation of usage volume and the significant computing power demand that Manus's architecture cannot currently generate. Additionally, users have been spoiled by various models, accustomed to getting results in minutes, and cannot tolerate waiting for hours.

Actually, if Manus can give users a predictable completion time, it may reduce the anxiety of waiting. After all, for a complex analysis task, even if the execution time is as long as 2-3 hours or even half a day, it's probably still much faster than a junior employee in reality.

Technical Capability

While computing power can be increased by adding more cards, the team's technical capability may be difficult to enhance in a short time. Without comparing it to a dream team like DeepSeek, Manus's technical reserves may be more suited for a user-end product with around one million daily active users.

A team of serial entrepreneurs may have an advantage in discovering needs and quickly creating a product ready for launch. But what if the user base grows larger? What if technical upgrades are needed? These are issues that need to be addressed over time. Currently, Manus users complain about the shell and slow running speed, with tasks lagging, which is an external manifestation of this problem (in the team's apology letter, we also saw that they didn't expect the product to become so popular).

Manus has an excellent starting foundation, with a founding team skilled at discovering and capturing user needs to guide the R&D direction. However, it still needs more AI talents to make Manus grow faster and more stable. After all, the current state of the product allows insiders to roughly understand the product architecture after using it a few times, and it wouldn't be difficult for large companies to copy it.

05 Manus Initially Proves Another Path for AI Agents

Agent products on the market have two directions: universal and vertically applied. The two most well-known products in the former category come from two large model vendors - OpenAI's "Operator" and Anthropic's "Computer Use." This direction can be considered the representative of universal Agents.

Anthropic's Agent software allows developers to complete various basic computer tasks such as input and file opening through a special API using the Claude model; OpenAI expands the range of hardware calls on this technology, allowing operations like programming, travel booking, and shopping on personal computers through API interfaces. AutoGLM from WisdomAI in China operates similarly to OpenAI's "Operator".

Vertical Agents are typically represented by Cursor and Devin. These two agent products have a good reputation among programmers but unfortunately have a narrow application field, making it difficult to promote them widely.

Manus's founder, Xiao Hong, also expressed in an interview that compared to vertical Agents, universal Agents undoubtedly have more universal value, but the former has faster application. In reality, many programmers are already using Cursor to write basic code.

Manus is positioned as a universal AI agent. Through engineering methods, it achieves Agent architecture innovation, encapsulating complex workflows on a cloud environment, integrating multiple large models, and automatically completing sub-node demand processing through task planning. The output is a complex task that would typically require multiple rounds of human-computer dialogue to complete.

Manus has achieved results that surpass OpenAI DeepSearch in some tasks, representing another viable path for Agents.

Universal agent products cannot rely solely on a single large model and are inherently the mission of third-party developers.

After all, it's highly unlikely that OpenAI and Anthropic's agents will be implemented through competing models, which determines that their product development progress and engineering methods cannot compare to those of third parties.

Simultaneously, Manus's application scope is more grounded compared to universal products and has a broader application field compared to vertical products. The series of operations promoted by universal agents, such as helping users purchase tickets, planning trips, and shopping on e-commerce platforms, pose no barriers for users to perform themselves. However, completing a Tesla financial report analysis report requires a high threshold. Products with thresholds usually have higher commercial value.

The internet community has always equated AI Agents with universal AGI, but the emergence of Manus tells us that even if AGI hasn't been achieved, who cares? For ordinary users, the one that can complete tasks is a good one.

06 Three Revelations

First, with AI assistance, independent decision-making ability is even more needed.

Even if Manus doesn't become a widely used agent product today or in the near future, we can still see a core issue from the continuously emerging similar products: independent thinking and innovation ability will be a person's core competencies in the future.

AI can indeed help people do many things, such as analyzing experimental results, writing experiment reports, and completing desk material writing. But AI won't help you conduct experiments, nor will it help you think about which point to start the analysis from to make the results look more impressive.

We have no doubt that with AI's involvement, the gap between people will become wider and wider.

Second, the source of first-hand information has changed.

Around 2010, during the last internet trend, it was primarily the major tech media that fueled the excitement. Beyond the four traditional internet portal websites, vertical tech media also played a crucial role. Prior to the AI trend, new startups often emerged from self-media, fermented on social media, and eventually gained widespread recognition.

In the "classical" era, maintaining contact with journalists ensured one stayed abreast of significant industry developments. However, in the AI era, investors often find themselves unprepared, unaware of the sudden rise of tech stars.

Many venture capital firms have established incubators to nurture their own startups, but the success rate remains less than ideal, with few blockbuster products emerging. How to seize potential or even emerging opportunities within a social media-driven model poses a new challenge for these institutions.

Thirdly, what should large companies do?

DeepSeek's exceptional model performance, coupled with its open-source advantage, has left companies like ByteDance, Tencent, and Baidu trailing in the technological wave. Manus has taken the initial step toward making universal agents a reality, once again leading the pack among large companies. Here, it would be unfair to compare Manus's R&D costs with those of large companies, as it would likely be embarrassing for the latter.

Should large companies follow suit or not?

References:

Zhang Xiaojun | Business Interview Series:

https://www.xiaoyuzhoufm.com/episode/67c3d80fb0167b8db9e3ec0f

https://pan.baidu.com/s/1KkGRdYdkf84vJygkFk2J5Q?pwd=n15q#list/path=%2F

Cyber Zen Mind:

Hands-on Review of Manus: The First Real AI Worker, Made in China (with 50 Use Cases + Breakdown)

Cyber Zen Mind:

Some Exclusive Information About Manus

Manus Usage Sharing