The largest financing of China's Coding Agent emerges, with Ant Group, Cathay Capital, Jinqiu, etc. making investments.
Text by | Zhou Xinyu
Edited by | Su Jianxun
In February 2025, Andrej Karpathy, a co - founder of OpenAI, proposed "Vibe Coding", which immediately became the most attractive track in AI entrepreneurship.
In this track that emphasizes "forgetting about the existence of code and creating through programming by conversing with AI", Lovable, the fastest - growing AI unicorn globally with an ARR (Annual Recurring Revenue) of $100 million, emerged.
In China, many well - known entrepreneurs have entered the Vibe Coding field. However, the hidden champion is a Shenzhen - based company, DeepWisdom.
You may not be very familiar with DeepWisdom, but several of the most well - known open - source projects in China in recent years originated from it:
The multi - agent framework project MetaGPT with nearly 60k stars on GitHub, and OpenManus, which was replicated by a team of 5 members in just 3 hours at night.
MetaGPT - X (abbreviated as "MGX"), a multi - agent Vibe Coding product launched by DeepWisdom in February 2025, achieved 500,000 registered users globally and an ARR of $1 million in just one month after its launch without any advertising investment.
Since its launch 7 months ago, MGX has maintained a stable growth rate. Official data shows that as of September 2025, MGX's monthly visits have reached 1.2 million, and the number of applications generated per day exceeds 10,000.
Wu Chenglin, the founder and CEO of DeepWisdom, revealed to us that MGX is currently the product with the largest user base in the Vibe Coding field in China.
Moreover, DeepWisdom has also secured the highest financing amount in the domestic Coding Agent track.
According to Intelligent Emergence, in the first half of 2025, DeepWisdom completed two rounds of financing from institutions such as Ant Group, Cathay Capital, Jinqiu Fund, BV Baidu Ventures, and Concept Fund, totaling approximately 220 million yuan.
Actually, before 2024, commercialization and impressive revenue figures were not the core goals of DeepWisdom.
Rather than a commercial company, DeepWisdom is more like a university laboratory. Internally, it encourages publishing papers, makes the core code accessible to almost all members, and has organized dozens of seminars on Self - play and Reward Model.
Wu Chenglin believes in the concept of academic cycle - only through continuous academic accumulation and breakthroughs can explosive success be achieved. For example, without a series of innovations such as MLA, the new MoE architecture, and self - play, DeepSeek would not have achieved the breakthrough of R1.
For the past two years, he has maintained a high - intensity habit of reading papers. During this period, he scanned nearly 200,000 papers on Arivix, carefully sorted out 2,100 of them, and marked less than 300 important studies.
"You must understand what's happening in the world and what really matters," Wu Chenglin told Intelligent Emergence. "As long as you focus on important things and follow the plan, it's not difficult for the product to achieve an ARR of $1 million."
For DeepWisdom, developing an AI Coding tool that can provide users with a complete commercial closed - loop is the most important thing at present.
Previously, Wu Chenglin had experience in large - scale and complex AI implementation projects with billions of users and trillions of data at companies like Huawei and Tencent. After founding DeepWisdom in 2019, he first worked on many B - end AI infrastructure and machine automation customization projects.
In the process, the path he discovered from the labor - intensive customization work is: use AI Coding to solve complete customization; further, the role of an AI Coding tool is not just that of an engineer, but a company that can help users achieve commercial monetization.
For this reason, from the open - source MetaGPT to MGX, DeepWisdom has designed a multi - agent framework that allows agents playing different roles such as researchers, product managers, and engineers to cooperate with each other, find a suitable SOP, and continuously self - feedback and iterate.
On January 13, 2026, DeepWisdom launched a new generation of MGX, which was renamed "Atoms".
Compared with its competitors, Wu Chenglin told us that unlike other products that can only generate "toys", Atoms has built - in systems for login, database, user authentication, deployment, and payment, and can deliver a complete website ready for online operation in just 5 minutes.
Another advantage lies in Atoms' high cost - effectiveness. Atoms can achieve more than 45% better results than market competitors with only 20% of the cost, making it more cost - effective than competitors like Lovable and Replit. "Users are still price - sensitive," Wu Chenglin said.
△ Comparison of cost - effectiveness between Atoms and its competitors. Source: DeepWisdom
AI Coding: From "Toys" to Helping Users Profit
Currently, the AI Coding track is booming, but most products are still criticized as "toys". The key lies in that the generated products cannot achieve end - to - end opening of the front - end and back - end, nor can they integrate third - party payment systems.
This results in most Coding applications generating only "preview versions" of single web pages, failing to truly improve efficiency and help users make profits.
Through built - in back - end, database, user authentication, and secure payment systems, Atoms develops products that can be directly launched for online operation.
Specifically, the multi - agent architecture enables Atoms to complete the entire product development and delivery process, including "requirement research - requirement document definition - prototype design - code development - data analysis and scraping" through natural language.
In a typical development process, Atoms will call on a research agent (for information collection), a product manager agent (for writing requirements and function comparison documents), an architect agent (for writing technical documents), an engineer agent (for website development), and a data analysis agent (for data scraping and analysis).
△ An online course platform built with Atoms, equipped with video courses, downloadable resources, and student progress tracking functions. Website: https://mgx - 2x5x65xztvp.mgx.world. Source: Company official
The advantage of multi - agent collaboration is also reflected in its ability to solve three bottlenecks of AI Coding products: in - depth research from market to user data, support for complex functions, and automated product testing feedback.
For example, to meet the in - depth research needs before and after product development, DeepWisdom has developed a research agent named "Iris", which can generate a complete research report in the form of summaries/audio, data charts, social media content/applications based on the user's research topic.
Iris. Source: Company official
In official evaluations, Iris' research ability surpasses that of Gemini - 2.5 - Pro, OpenAI, Kimi, and Perplexity.
△ Evaluation of Atoms - DeepResearch's capabilities. Source: DeepWisdom
In addition to solving the product payment problem by integrating third - party payment capabilities such as Stripe, Atoms also has a built - in SEO Agent named "Sarah". Simply put, Sarah is an SEO expert. By constructing an automated SEO strategy for the products built by users, Sarah can help the products obtain organic traffic from search engines.
Therefore, the products built through Atoms can be directly launched for online operation and achieve profit and automated growth.
△ The process of building a jewelry e - commerce website with Atoms. Source: Company official
Within DeepWisdom, academic research, open - source, and closed - source commercialization form a closed - loop around agents.
In Wu Chenglin's view, academic research helps the team determine the iteration direction of products and technologies.
So far, DeepWisdom has submitted 9 papers to NeurIPS, the top conference in the computer field, and 5 of them have been accepted. In the most prestigious Oral (presentation) session, 3 papers from DeepWisdom have been selected.
Through academic research, DeepWisdom has developed an approach for building the next - generation agents:
Through the dynamic division of labor and reasonable routing of multi - agents, form a suitable SOP like the human project management process, and gradually improve key capabilities such as self - evaluation, memory management, and cross - environment operation.
Later, both the first - generation commercial product MGX launched by DeepWisdom and the popular OpenManus were realized based on the concept of multi - agent collaboration.
The currently launched product Atoms also operates under the multi - agent framework, where a multi - agent team collaborates to complete the entire product operation process, including market research, full - stack development (including user authentication, payment, and database back - end capabilities), deployment and operation, product marketing, and data analysis.
△ Atoms calls on 7 agents with different roles during the product development process. Source: Atoms official website
The path to quickly verify technologies and products and gain inspiration is open - source.
Early on, in June 2023, DeepWisdom open - sourced the agent framework MetaGPT, which initially verified the feasibility of multi - agent collaboration and SOP management in programming. After its release, this project not only gained 58.8k stars on GitHub but also brought tens of thousands of new members to the DeepWidsom community.
What's less known is that almost a year before Manus, Wu Chenglin and Lin Junyang, the technical leader of Alibaba Tongyi Qianwen, promoted an open - source Coding Agent project named OpenDevin, which is the open - source framework on which the later OpenHands, with over 60,000 stars on GitHub, is based.
Based on this project, several interns at DeepWisdom were able to replicate Manus in just 3 hours at night.
Currently, DeepWisdom's open - source organization, Foundation Agents, has more than 150,000 stars on GitHub.
The success in open - source has also established a good technical reputation for DeepWisdom's commercialization.
The commercial product MGX was launched in February 2025 at a rather unfavorable time.
Wu Chenglin recalled that at that time, the company's financing had not been secured, the cash flow was tight, and even the servers and large - language models were obtained on credit from UCloud and AWS. There was almost no budget for MGX's promotion, and it relied entirely on the members' editing of promotional videos and the community members' forwarding.
However, the previously accumulated reputation led to an unexpectedly good growth effect for MGX - in just one month after its launch without any advertising investment, MGX's ARR (Annual Recurring Revenue) from subscriptions exceeded $1 million. Wu Chenglin revealed to us that MGX's ARR has been steadily increasing since its launch one year ago.
However, to catch up with Lovable's ARR of $100 million, MGX must achieve cross - boundary growth.
Wu Chenglin made two key decisions: one is renaming, and the other is optimizing the product's cost - effectiveness.
Renaming a product with more than 500,000 users undoubtedly poses a huge risk. However, Wu Chenglin believes that in the English context, the pronunciation of MGX is not user - friendly for ordinary users, and renaming it to "Atoms" is more conducive to C - end dissemination.
As for the iteration direction of high cost - effectiveness, it stems from the most basic product concept: a product that can capture users' hearts should first help users make money and second, be efficient and cost - effective.
Official test results show that Atoms can achieve better results than competitors like Lovable and Replit at any cost level.
The high cost - effectiveness is achieved through the optimization of the combined effects of multiple open - source models such as DeepSeek and Qwen. Wu Chenglin told us that this proves that at present, closed - source models cannot establish an absolute advantage over open - source models. "The effect of Claude is not the best."
△ Atoms outperforms its competitors in performance at the same cost. Source: DeepWisdom
By creating a high - cost - effective and "one - stop" development process, Atoms enables more ordinary people, such as "programming beginners", to unleash their development and business operation needs and obtain rewards from them.
An Atoms user who impressed the DeepWisdom team is a Canadian mechanic. This ordinary blue - collar worker with no programming background developed a 2D robot battle game with a plot on Atoms using his mobile phone after work.
Now, he has established a game community of 15 - 20 test players on Discord, continuously collects feedback, and iterates the game.
