By Wang Haonan, People’s Daily
Near midnight, a meeting room in Beijing’s Haidian district was still lively with discussion. On a whiteboard, two Chinese characters — “speed” and “accuracy” — were circled in bold.
For Wang Huan, an algorithm researcher at Moonshot AI, and his colleagues, such late-night debates have become routine as they tackle one of the latest challenges in artificial intelligence: multi-agent systems.
Comparisons of benchmark test results evaluating the coding capabilities of major AI models show that Kimi K3 sits at the leading edge. (Photo courtesy of Moonshot AI)
What is a multi-agent system?
“In the past, a single AI agent was like a capable assistant,” Wang explained. “Now we want it to work more like a project manager that can coordinate an entire team.”
Wang offered a simple analogy.
“Imagine a family preparing for a move. One person packs the books, another arranges transportation, someone else checks the inventory, and finally everything is reviewed together. A multi-agent system works in a similar way. A lead agent first breaks down a complex task, then coordinates multiple specialized sub-agents to work on different parts simultaneously.”
AI has become a key driver of the latest wave of technological revolution and industrial transformation.
“Multi-agent systems represent a new frontier for scaling AI models,” said Yang Zhilin, founder of Moonshot AI. “The rapid pace of technological progress reminds us that continuous innovating is essential.”
In January this year, Moonshot AI released and open-sourced its Kimi K2.5 model, introducing multi-agent capabilities for the first time. The upgrade significantly reduced the number of steps required to complete complex tasks.
Three months later, the company launched Kimi K2.6, capable of coordinating 300 sub-agents to carry out 4,000 collaborative task steps in parallel. Under an autonomous agent framework, it can operate independently for up to five consecutive days.
On July 16, the company unveiled a more intelligent model, Kimi K3. With 2.8 trillion parameters, it is currently the world’s largest open-source AI model by parameter count.
A stronger technological foundation is opening new opportunities for industry growth. By mid-June this year, Moonshot AI’s annual recurring revenue had tripled compared to March. More than 70 percent of its revenue now comes from application programming interface (API) services.
“Our number of paying overseas users has quadrupled, and our products are now available in more than 200 countries and regions,” said Huang Zhenxin, the company’s head of business operations.
Beijing is accelerating efforts to position itself as a global center for AI innovation. The city is rolling out nine major initiatives covering areas such as original technology innovation, strengthening an independent intelligent computing ecosystem, and improving high-quality data resources.
“We will continue to leverage Beijing’s strong concentration of talent while improving policy support,” said Lin Jianhua, deputy director of the Beijing Municipal Commission of Development and Reform. “By focusing on core technology breakthroughs, shared technology platforms and demonstration applications, we will launch major projects to foster new forms of the intelligent economy.”
During the first half of this year, Beijing added 22,000 PFLOPS of new computing capacity, bringing the city’s total computing power to 82,000 PFLOPS.
Across Beijing, AI startups are also growing rapidly.
Beijing-based AI infrastructure platform SiliconFlow, which aims to build a “token factory for the AI era,” has gained more than 10 million registered users.
Meanwhile, ShengShu Technology, another company in Beijing which focuses on developing general-purpose world models, now serves more than 40 million individual users and over 10,000 enterprise users across more than 200 countries and regions.
Looking ahead, Moonshot AI plans to further expand its open-source strategy.
“In the second half of the year, we will continue investing in more efficient model architectures, longer context windows and more capable multi-agent systems,” Yang said. “Through solid foundational technological innovation and an open approach to collaboration, we hope to make more AI technologies widely accessible and benefit more users.”
