In February 2025, when Andrej Karpathy put a name to an emerging idea, the phrase “vibe coding” quickly entered the startup lexicon. The concept was straightforward: ignore the mechanics of programming and build software by interacting directly with artificial intelligence.
That idea soon translated into commercial momentum. Lovable, a startup built explicitly around vibe coding, reached USD 100 million in annual recurring revenue in record time, according to the company.
In China, the push into vibe coding has attracted a wave of high-profile founders. The category’s quiet leader, however, has emerged from a Shenzhen-based company that remains relatively obscure outside developer circles: DeepWisdom.
The name may be unfamiliar to mainstream audiences, but DeepWisdom sits behind several of China’s most widely used open-source AI projects in recent years. Among them is MetaGPT, a multi-agent framework that has accumulated close to 60,000 stars on GitHub. Another is OpenManus, which the company says five team members recreated in roughly three hours late one night.
The same month Karpathy popularized the term vibe coding, DeepWisdom released its own multi-agent product, MetaGPT-X, or MGX. Without paid marketing, the company said MGX reached about 500,000 registered users globally within a month and crossed USD 1 million in annual recurring revenue. Seven months later, growth had continued. According to DeepWisdom, by September 2025 the product was drawing roughly 1.2 million monthly visits and generating more than 10,000 applications per day.
Founder and CEO Wu Chenglin said MGX has become the largest vibe coding product by user base in China. The company has also raised more capital than any other domestic player in the coding agent space. In the first half of 2025, DeepWisdom completed two funding rounds backed by Ant Group, Cathay Capital, Jinqiu Capital, MindWorks Capital, Baidu Ventures, and Concept Capital, raising roughly RMB 220 million (USD 30.8 million).
Research before revenue
Until recently, revenue metrics were not DeepWisdom’s primary focus. The company often operated more like a university lab than a conventional startup. Team members were encouraged to publish academic papers, core code was broadly accessible internally, and the company hosted frequent seminars on topics such as self-play and reward models.
Wu has argued consistently that commercial outcomes follow sustained research. He often points to DeepSeek as an example, saying that earlier work on mixture-of-experts architectures and self-play enabled later breakthroughs such as R1.
For the past two years, Wu has maintained an intense reading schedule. He estimates that he has skimmed nearly 200,000 papers on arXiv, closely reviewed about 2,100, and identified fewer than 300 as genuinely consequential.
“You have to understand what’s happening in the world, and what actually matters,” Wu said in an interview with 36Kr. “If you focus on the important things and follow your plan, getting to USD 1 million in ARR isn’t hard.”
DeepWisdom’s current priority is building AI-driven coding tools that allow users to move from idea to commercialization, not simply generate isolated code snippets.
Before founding the company in 2019, Wu worked at Huawei and Tencent, leading large-scale AI deployments. DeepWisdom initially focused on enterprise AI infrastructure and automation projects, experiences that shaped Wu’s views on software development.
His conclusion was blunt: AI coding tools should eliminate customization. More importantly, they should not behave like individual engineers, but like companies that help users commercialize ideas.
That philosophy underpins DeepWisdom’s multi-agent architecture, used across both MetaGPT and MGX. Different agents assume roles such as researcher, product manager, and engineer, working through structured workflows that evaluate and refine outputs continuously.
From MGX to Atoms
On January 13, 2026, DeepWisdom released a new generation of MGX and rebranded it as Atoms.
Wu said Atoms differs from competing products by targeting commercial deployment rather than hobby projects. It includes built-in systems for login, databases, authentication, deployment, and payments, allowing users to launch functional websites quickly.
Atoms also competes on cost. Wu said the product delivers better results than competitors while costing significantly less, outperforming products such as Lovable and Replit on price-performance. This is a company claim and has not been independently verified. “Users are still price-sensitive,” Wu said.
AI coding tools are proliferating, but many are criticized for stopping at surface-level outputs. Most generated applications remain previews: single-page frontends without backend integration, open APIs, or payment systems, limiting their commercial usefulness.
Atoms takes a different approach. By embedding backend infrastructure, databases, authentication, and secure payments, it produces applications that can go live immediately. Its multi-agent architecture supports an end-to-end workflow that moves from natural language prompts through research, requirements definition, prototyping, development, and analysis.
To support research-heavy use cases, DeepWisdom built a dedicated research agent called Iris. Based on a user’s topic, Iris generates reports in formats ranging from summaries and charts to audio and social media-ready content. In internal benchmarks, DeepWisdom said Iris outperformed Gemini 2.5 Pro, OpenAI models, Kimi, and Perplexity on research tasks. These results have not been independently validated.
Atoms also integrates third-party payments such as Stripe and includes an SEO-focused agent named Sarah, which automatically generates search optimization strategies for products built on the platform.
Open source as a proving ground
Inside DeepWisdom, academic research, open source, and commercialization form a feedback loop. The company has submitted nine papers to NeurIPS, five of which were accepted, including three selected for oral presentation.
That research informed the company’s work on next-generation agents, including dynamic role assignment, intelligent task routing, and the integration of self-evaluation and memory management.
Open source has served as a testing ground. In June 2023, DeepWisdom released MetaGPT, demonstrating that SOP-based multi-agent collaboration was viable. The project has since accumulated about 58,800 GitHub stars.
Less widely known is that nearly a year before Manus attracted attention, Wu and Lin Junyang, a technical lead on Alibaba’s Qwen project, had already advanced an open-source coding agent framework called OpenDevin. That work later underpinned OpenHands, which has surpassed 60,000 stars on GitHub. The same foundation allowed several DeepWisdom interns to replicate Manus in a matter of hours.
Today, DeepWisdom’s open-source organization, Foundation Agents, has accumulated more than 150,000 GitHub stars.
Betting on scale
MGX launched under financial strain. Funding had not yet closed, cash flow was limited, and the company relied on server and model credits from UCloud and Amazon Web Services. Marketing consisted largely of short videos edited by employees and community sharing.
MGX nonetheless crossed USD 1 million in ARR through subscriptions alone, according to Wu, with revenue continuing to grow. Closing the gap with Lovable, however, required expansion beyond its existing user base. Wu responded by rebranding MGX as Atoms and emphasizing cost efficiency.
The rename carried risk. Wu said “MGX” was difficult to pronounce in English and poorly suited for consumer adoption. “Atoms,” he argued, would travel more easily.
Cost efficiency remains central. According to internal evaluations, Atoms outperforms Lovable and Replit at comparable price points by combining multiple open-source models, including DeepSeek and Qwen, rather than relying exclusively on closed systems. Wu said closed models do not yet offer a decisive advantage.
Atoms has attracted users beyond traditional developer circles. One example frequently cited internally is a Canadian car mechanic with no programming background who built a story-driven 2D robot battle game using only his phone. He later formed a small Discord community of testers and continues to iterate based on feedback.
Not a one-person company
For Wu, the more interesting question is organizational design in the AI era. DeepWisdom takes a contrarian stance, arguing that one-person companies and very small teams are unlikely to hold an advantage as competition intensifies.
Examples such as Midjourney, Cal AI, and OpenArt have popularized the idea that small teams can scale dramatically with AI. Wu disagrees. “When everyone has computers, it’s as if no one does,” he said at the 2025 Inclusion Conference. “AI doesn’t reduce competition. It intensifies it.”
He pointed to Lovable’s rapid headcount growth, from 15 employees in late 2024 to 45 by June 2025, as evidence that speed still requires manpower. DeepWisdom now employs more than 80 people and expects to reach 100–120 by the end of 2025.
Wu attributes efficiency to management, not headcount. In past roles, he observed that rigid divisions of labor created hidden costs. “Some companies spend eight hours arguing during the day and only write code for two hours at night,” he said.
He cited Anthropic as an example of how AI can reduce organizational friction. Employees there use Claude Code as a persistent collaborator to summarize work and suggest improvements, forming rapid feedback loops.
Over time, Wu expects AI to act as a de facto manager, assigning tasks based on employees’ skills and working styles. Until then, DeepWisdom prioritizes generalists. Fewer than 15 core employees operate across frontend, backend, algorithms, and testing to minimize handoffs.
Many hires have come from the company’s user community and internal hackathons rather than traditional recruiting. Wu believes performance under real constraints reveals more than resumes.
DeepWisdom is preparing to open its first overseas office in Silicon Valley, in a building informally known as “the lucky office,” which has previously housed companies such as Google, PayPal, and Logitech. The building also happens to share a wall with Stanford University.
“For a long time to come,” Wu said, “the top tier of AI companies will still be competing on human efficiency.”
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.