Hi, this is Tony! Welcome to this issue of Recode China AI (for the week of March 3, 2025), your go-to newsletter for the latest AI news and research in China.

Personal Update: I left Baidu last month and am joining a new company (unrelated to AI) this week, so updates to this newsletter may become less frequent. I’ll continue monitoring AI developments in China but will highlight only the most impactful stories. Thanks in advance for your understanding! This newsletter is very important to me, and I’ll keep delivering quality content.

Over the past week, an AI agent called Manus quickly gained popularity across the Chinese tech community. It is hailed as the GPT or DeepSeek moment in AI agent, a type of software program that can autonomously perform tasks on behalf of users.

Manus can handle a wide range of tasks such as producing a personalized travel handbook, creating an interactive course on the scale of universe, analyzing Tesla’s stock trends. It has also achieved (state-of-the-art) SOTA performance on the GAIA benchmark, surpassing OpenAI’s Deep Research.

However, it also sparked controversy, as critics dismissed the product as merely a “wrapper” around existing open-source technologies. The underlying models of Manus include Anthropic’s Claude and fine-tuned Alibaba’s Qwen. Additionally, two similar projects were open-sourced within days.

What is Manus? How good is it? Who’s behind Manus? Let’s dive in.

What’s New: On March 6, Beijing-based AI startup Monica.im (the company’s Chinese name is 蝴蝶效应, translated as Butterfly Effect) released Manus, touted as the first general AI agent.

Named after the Latin word for “hand,” Manus goes beyond traditional AI chatbots by actively performing tasks using a sophisticated Multiple-Agent architecture. According to the company, each agent operates independently within a virtual machine environment, capable of breaking down queries into to-do steps, writing and executing code, accessing web resources, and interacting with applications autonomously.

While Manus has yet to launch publicly, it has provided early testing access to invited users. On one Chinese second-hand marketplace, invitations were reportedly selling for tens of thousands of yuan. The product has also started gaining traction on X.

How it Works: Manus showcases its prowess through diverse practical demonstrations:

• Resume Screening: Manus automatically unpacks compressed files, reviews candidate resumes, ranks applicants based on their expertise, and outputs data into structured Excel tables.

• Real Estate Selection: Manus assists users seeking properties in New York by analyzing neighborhood safety, school quality, and affordability. It executes Python scripts to calculate budgets based on user income and filters suitable properties, compiling findings into professional, actionable reports.

• Stock Market Analysis: Manus retrieves historical stock data (e.g., NVIDIA, Marvell, TSMC), verifies accuracy through multiple sources, and uses Python-driven visualizations to clearly illustrate market correlations and insights.

On the GAIA benchmark—evaluating real-world task-solving capabilities—Manus set new SOTA records across all levels.

Currently, mainstream AI agent services remain costly—ChatGPT Operator charges $200 per month, while developer-focused AI service Devin costs $500 per month. In contrast, Manus is free during its testing phase, with an estimated per-task cost of $2.

Hands-On Testing: I don’t have an invitation code, so I relied on firsthand reviews from Lanxi, a well-known tech influencer in China.

One standout example is Manus created an interactive web-based game within an hour, allowing users to role-play as Google’s CEO, making key historical decisions to learn about company culture. Other tests include:

Personal Homepage Creation: Manus swiftly designed a Linktree-style personal webpage, autonomously gathering user data online and coding a functional, interactive page.

Engineering Assistance: Manus independently accessed documentation from Atlas’ official site to troubleshoot a robotic arm, bypassing traditional support routes and saving expenses.

Historical Timelines: When tasked with crafting a minimalist history timeline for a country, Manus designed a visually engaging, interactive web-based timeline segmented by historical eras.

Game Development: Manus quickly developed a match-three puzzle game using tech company logos. Although initially hindered by download restrictions on third-party websites, adjusting the requirements allowed Manus to autonomously complete the task.

Limitations: Manus still struggles with complex tasks. For example, screen adaptability issues require clearer instructions, and while Manus responds quickly to modifications, server outages have occasionally disrupted progress.

Manus typically takes more than 30 minutes and even hours to complete a task, so the user experience feels slow. Considering that completing the same task manually would likely take even more time, using an AI agent with this level of performance to assist with work should be a good choice.

Behind Manus: Behind Manus are Monica.im Founder and CEO Xiao Hong (Red), Co-founder and Chief Scientist Ji Yichao (Peak), and Product Partner Zhang Tao (hidecloud)—two entrepreneurs born in the 1990s and a seasoned product veteran born in the 1980s who has navigated 10 companies in 15 years.

(Update) According to public information, Monica.im, founded by Xiao Hong in 2022, is the company behind Manus. Manus AI operates as a separate project within Monica.im and is led by Xiao Hong, Peak Ji, and Zhang Tao. On X, Peak Ji lists his title as Co-founder of Manus AI.

This small, scrappy team unexpectedly created a notable storm using garage-startup tactics.

Deep tech founders typically have prestigious backgrounds, while consumer-product founders usually emerge from humble beginnings, says Liu Yuan, an investor from ZhenFund and an angel investor in Monica.im.

Xiao Hong, from Ji’an, Jiangxi Province, is a graduate of Huazhong University of Science and Technology (HUST). He began his entrepreneurial journey in Wuhan starting 2016 by developing two successful WeChat plugins—Yiban and Weiban—which were later acquired by a unicorn company.

In 2023, Xiao and his AI company, Monica.im, launched Monica, a Chrome extension powered by the ChatGPT API for chatting and copywriting. The product achieved an ARR of $10 million in its first year and reached 10 million users by 2024.

Peak Ji, a Beijing native, began entrepreneurship in high school. Shortly after Apple launched its App Store, he developed Mammoth Browser and later Magi, a knowledge search engine.

Zhang Tao, originally from Chongqing, has held product roles across diverse companies such as Tencent, Wandoujia, ByteDance, and others spanning B2B and B2C markets.

All three had previous ties to ZhenFund, a Beijing-based venture capital firm. Both Xiao and Ji’s first companies received ZhenFund backing before being acquired, while Zhang’s former employer, Beyond Lightyears, was also ZhenFund-funded. The trio partnered in 2024 through ZhenFund’s introductions.

Claude Wrapper: The sudden surge of attention has sparked controversy. One X user, through jailbreaking, discovered that Manus is Anthropic’s Claude equipped with 29 tools. It also leverages an open-source project called Browser-Use, which enables AI agents to access websites.

In response, Peak Ji said Manus integrates open-source tech, including Browse-Use, and plans to release more open-source contributions soon. Manus uses Claude and fine-tuned Qwen LLMs and is testing the latest Claude 3.7 internally.

It remains unclear whether Manus introduces any internal technological innovations or is merely an impressive UI wrapper.

I’m not dismissing the idea of a wrapper—as long as the product delivers real value to users at an affordable cost. The key question is if it’s merely a wrapper, competitors could quickly replicate and surpass it.

Open-Source Implementations: Meanwhile, Chinese open-source platform CAMEL AI quickly released OWL, an open-source alternative positioned as a fully autonomous multi-agent productivity tool.

OWL achieved a performance score of 57.7% on the GAIA Benchmark, surpassing existing records and sparking enthusiasm within the open-source community. Their project is now available on GitHub.

Another open-source alternative is called OpenManus, allowing users to deploy their own AI agents locally. Unlike Manus’ cloud-based approach, OpenManus directly operates users’ personal computers. Its development team, MetaGPT, outlined future improvements, including enhanced planning tools, real-time demonstrations, replay capabilities, reinforcement learning models, and comprehensive benchmarks.

Why It Matters: Manus’s wild popularity has brought AI agents into the spotlight in China. As more companies jump on the bandwagon and models become increasingly capable in reasoning, it signals a shift in expectations—from AI copilots to full-fledged digital agents.