Manus: Is China’s overnight AI sensation a breakthrough or just hype?

In early March, general-purpose artificial intelligence (AI) agent Manus, created by AI startup Butterfly Effect, took the Chinese internet by storm within 24 hours of its launch, becoming a phenomenon in the wake of DeepSeek.

Its explosive popularity stems from its positioning as the “first general-purpose AI agent” capable of autonomously breaking down and executing complex tasks (such as screening resumes, analysing stocks, and generating PowerPoint presentations), delivering tangible results directly rather than merely providing textual suggestions.

Its controlled access invitation code mechanism creates a sense of scarcity and has further fuelled the hype, with prices on the second-hand market reaching as high as 50,000 RMB (roughly US$6,914). The official website even crashed multiple times due to the traffic spike.

However, Manus’s popularity has been controversial from the start, primarily due to three aspects. One, the AI agent was accused of “encapsulating” Anthropic’s Claude and Alibaba’s Qwen, lacking independent research and development capabilities for underlying models, raising doubts about its core technological advantage.

Two, the actual user experience falls short of what is depicted in promotional videos, with some testers reporting issues such as lag, repetition and hallucination. 

Three, marketing concerns. Manus has been criticised for overreliance on social media hype and Chinese internet dissemination, leading to accusations of being all talk and no action. Some of its demonstrated functionalities have also been accused of merely “stitching together” existing technologies. 

The controversy stems from the fact that Manus is not an underlying technology but an agent that integrates other AI models. Its goal is not to push the boundaries of AI capabilities, but to integrate different AIs and make their use more efficient. For example, Manus is more similar to a product like AI leader OpenAI’s Deep Research.

I think that many tech media outlets are making a flawed assessment by dismissing Manus’ value simply because it does not offer groundbreaking innovation at the large language model (LLM) level. Historically, technology integration is also a crucial — sometimes, the most crucial — part of technological revolution. Let us not forget that the catalyst for the last mobile internet wave, Apple’s iPhone, did not feature any fundamental innovations in its underlying technologies either — it was itself a product of technology integration as well.

Take resume analysis as an example — both OpenAI and DeepSeek offer this capability, but widespread adoption in HR departments will only happen as specialised products like Manus mature and refine their solutions.

If we compare OpenAI and DeepSeek to the technologies that forge a blade, widespread adoption requires more than just the raw material. Like crafting a knife, we need to design the shape of the blade, the handle, the sheath, and other components to create a complete, usable product. Only then can this “knife” permeate all aspects of economic life.

While Manus may not represent a groundbreaking innovation or breakthrough in the underlying technology of LLMs, it does not mean that it lacks advancements in engineering. Its “multi-agent architecture” is like a well-coordinated team with three important upgrades: the ability to break big tasks into smaller steps (planning), knowing which tools to use for each job (tool use), and remembering past interactions (memory).

The real game changer here is the planning ability — it’s like teaching AI to take a complex project and automatically figure out all the steps needed to complete it.

The AI behind Manus builds on technology first shared publicly last October by one of its creators, Ji Yichao. By integrating the QwQ-32B LLM, it achieves two key innovations: one, a “long chain-of-thought reasoning” — instead of jumping straight to answers, the AI breaks down problems step-by-step (like showing your work in math class); two, action planning — it doesn’t just think — it executes tasks in a logical order, much like following a recipe. These features power Manus today, helping it tackle complex jobs methodically.

Notably, the QwQ-32B LLM, which was released in the same time period as Manus, has already shown impressive advancements in AI agent capabilities. Analysts believe that adopting a fine-tuned QwQ-32B model could significantly boost planning accuracy and further enhance intelligent agent products.

The tech world is now laser-focused on improving the planning abilities of underlying AI models, a key factor in how well intelligent agents perform. These advancements will directly shape the industry’s competitive landscape.

The scepticism that Manus has faced upon its release from many Chinese media outlets is actually a “tradition” within the Chinese tech media community. When DeepSeek first gained prominence in the AI technology community through its price advantage, most Chinese tech media outlets either paid little attention or dismissed this cost advantage.

The differing responses to technology integration efforts, such as the widespread recognition of OpenAI’s Deep Research in the US and the scepticism faced by Manus in China, highlight the fundamental differences between the two tech media ecosystems.

US tech media has very close and open interactions with academic and industry groups at top US universities. Top university research teams, academic groups, and industry companies in the US operate almost as a unified entity, so that the US tech media circle can more accurately grasp the pulse of technological trends.

Unlike the US’s relatively relaxed technological innovation environment, due to various reasons, China’s emerging tech companies often face intense commercial battles and competition from the outset, while also navigating a more stringent financing environment. As a result, they are more cautious and secretive about their development plan and core technologies, which leads them to intentionally maintain a certain distance from the media. Thus, China’s tech media has long been pessimistic and sceptical about China’s technological development.

Related: Why DeepSeek chose text over images: The untold strategy behind its AI success | DeepSeek: Changing the game, not the leader