China may crack down on the acquisition on Manus as it may consider it nationally important technology, or something along those lines. We will see in the coming weeks.
Below is a translated version of an article on this topic posted on WeChat yesterday, titled Farewell, Large Models; Hello, Agents: Understanding the Paradigm Shift Behind Meta’s Acquisition of Manus (别了，大模型；你好，Agent：读懂Meta收购Manus的范式转移), authored by TechAstra星科技.
I’m sharing the article below without further commentary.
On the second-to-last day of 2025, a thunderclap came from across the ocean.
Meta (formerly Facebook) announced that it would acquire AI startup Manus outright for several billion dollars. This is Mark Zuckerberg’s third-largest deal in Meta’s history, after the $19 billion acquisition of WhatsApp and its stake in Scale AI.
As the news dominated the tech world, another, more hidden story was unfolding in parallel: months earlier, Manus’s Beijing office had already been emptied out. The Chinese engineers who once helped this butterfly take flight, holding their N+3 severance packages, watched as the company fully severed its ties to its homeland and turned around to become a jewel in the crown of a Silicon Valley giant.
It is a victory with a distinctly magic-realist hue.
Manus—born in Wuhan, China; grown in Beijing; and ultimately “ascending” via a Singapore shell—used just three years to travel a path many tech companies cannot complete in a lifetime. It is a typical story of struggle for a new generation of Chinese technical elites, and also a modern apocalyptic parable about survival, choices, and costs amid the cracks of global geopolitics.
What Manus leaves behind for China’s tech community is not merely a legend of “getting rich by going abroad,” but a modern fable loaded with metaphor: amid a complex chessboard of compute blockades, capital decoupling, and technological paradigm shifts, what posture must China’s sharpest minds adopt to take a seat at the world’s table?
To understand why Manus is worth several billion dollars, we cannot look only at AI—we must first look at the people.
Manus’s core team is not the typical cohort of academy-trained AI scientists. Their most prominent labels are extreme engineering capability and a keen grasp of traffic dynamics and human nature.
The founder, Xiao Hong, is an underestimated operator. As early as 2015, he founded Nightingale Technology and developed “Yiban” and “Weiban.” If you have ever worked as a WeChat public-account editor, there is a high chance you have used their plugins.
Even then, Xiao had already internalized a principle: inside a giant’s ecosystem (WeChat’s), building an extremely useful “add-on” can be highly lucrative.
A decade later, he transplanted that logic into the AI era. Only this time, the giant shifted from Tencent to OpenAI and Anthropic, and the “add-on” evolved from a browser plugin into an AI agent.
But the person who truly gave Manus its technical soul was its chief scientist, Ji Yichao (Peak Ji).
To veterans of China’s hacker community, the name “Ji Yichao” represents a legend. In 2011, while still in high school, he single-handedly wrote the Mammoth browser, stunning Sequoia Capital and Xu Xiaoping. He is a classic lone-wolf genius, with an almost obsessive pursuit of human-computer interaction.
Before joining Manus, Ji Yichao tried to train an NLP model from scratch. But he quickly ran into the famous bitter lesson: the model he painstakingly trained for months was, overnight, reduced to scrap metal by GPT-3’s dimensionality-reduction-style superiority.
That blow led him to a new technical philosophy—and became the foundation of Manus: orthogonality.
Ji wrote on his blog: “If progress in the underlying models is a rising tide, what we should build are boats on the surface, not pillars planted in the riverbed.”
In other words, Manus would never compete with OpenAI on model parameters. Instead, it would focus on how to use models better. This technical path—often called context engineering—allowed Manus to avoid a direct confrontation with the giants. Instead, it leveraged the giants’ capabilities to build products stronger than the giants’ own.
Why would Meta spend several billion dollars to buy a company that does not build its own models?
Because the wind direction in AI has changed. Over the past two years, we have been immersed in ChatGPT’s dialog box, marveling that it can write poetry and code. But the business world quickly discovered an awkward reality: chatting does not directly create productivity. A boss does not need a robot that can keep you company. A boss needs an employee who can finish the work.
Manus seized precisely this pain point. It is not a chatbot. It is an agent.
On the authoritative GAIA benchmark, Manus’s score crushed OpenAI’s Deep Research. What can it do?
You tell it: “Help me research the Southeast Asian coffee market in 2025.”
It will not merely output a block of text. Like a real human employee, it will: open a browser; search dozens of web pages; read PDF financial reports; filter out ads and junk; fill the data into Excel; and finally produce a PowerPoint and send it to you.
To achieve this, Manus did not “refine” some earth-shattering large model. Instead, it built a virtual operating system.
They equipped the AI with a virtual file system as an external brain, solving the problem of large models failing to retain long context. They designed a deterministic state machine to prevent the AI from hallucinating and randomly clicking the mouse while working.
This is a classic victory of engineering. While Silicon Valley was still enthralled by the Scaling Law, this group of Chinese engineers—through extreme optimization and deep understanding of user scenarios—proved that application-layer innovation can be worth a fortune as well.
Yet the more successful Manus’s technology became, the more awkward its position grew.
In July 2025, Butterfly Effect, Manus’s parent company, made a decision that sparked enormous controversy at the time: it moved its global headquarters from Beijing to Singapore and carried out large-scale layoffs of its China team.
On social media, criticism was widespread. Some called it a “backstab.” Others said, “They ran the moment they were fully grown.” But if we cut through the emotional fog, we see a chilling commercial logic.
In front of Xiao Hong stood an unsolvable “impossible triangle”:
Compute: To train and run top-tier agents, you must rely on Nvidia H100/H200 clusters. But on the U.S. Commerce Department’s export control list, those chips cannot legally enter China.
Capital: Manus needed U.S. dollar funds to sustain expensive compute burn. Its Series B was led by top Silicon Valley VC Benchmark, but under the Biden administration’s AI investment ban, the condition for taking that money was: you cannot be a Chinese company.
Ecosystem: Manus depends underneath on Claude and GPT APIs. If it retained a China identity, it could face the risk of being “cut off” by upstream providers at any time.
To survive, it had to leave.
It was a migration like amputating an arm to save a life. More than 40 key technical pillars were transferred overnight to Singapore; the remaining 120-plus ordinary employees were dissolved on the spot under N+3 or even 2N compensation plans.
At that moment, Manus completed a de-China-ization in both physical and legal terms. It became a Singapore company, taking American money, using American chips, and serving global customers.
This may be the most helpless template for China’s top hard-tech entrepreneurs over the next decade: “Chinese brains + a Singapore shell + American capital + a global market.”
Zuckerberg has calculated this very carefully.
Meta has the strongest open-source model in the open-source world—Llama—but at the application layer it has long lacked a killer product. It has a brain, but it lacks a pair of nimble hands.
Acquiring Manus instantly fills Meta’s gap in agentic AI. Imagine this: in the future, within WhatsApp, there will not only be a Meta AI that can chat with you, but also a Manus that can directly help you book tickets, buy things, and handle work. This is not just technical integration—it is the closing of a commercial loop.
For China’s AI industry, however, it brings an inexpressible sense of loss.
We watch a group of exceptionally talented young Chinese people, compelled by the gravity of geopolitics, forced to graft the fruits of their intelligence onto someone else’s tree. Manus proved that Chinese engineers are fully capable of defining the next generation of AI products. They are no longer “Copy from US.” They produced “Original from China” innovation.
But that innovation cannot nourish the domestic ecosystem.
Our shortage of compute has driven away the companies that most need compute.
Our capital environment cannot retain the unicorns that most need to burn money.
Our market isolation prevents the best products from serving the users they understand best.
Manus is like a butterfly crossing the vast sea. On the far shore, it stirred a storm worth billions of dollars, proving to the world the rise of Chinese technological strength. But when it looked back, the body that once nurtured it could only watch its receding silhouette and let out a complicated sigh.
In any case, we should congratulate Xiao Hong, Ji Yichao, and the Manus team.
In the cracks of history, through extreme technology and decisive courage, they carved out a road of their own.
And for entrepreneurs still holding the line at home, Manus’s story is both a lighthouse and an alarm bell. In the next cycle, how to find one’s place in this fractured world will be a harder problem to solve than writing code.