At a roundtable discussion at the ZGC Forum, Luo Fuli, head of large models at Xiaomi, dropped a bombshell on the industry: Artificial General Intelligence (AGI) has already been achieved, and the sector's next goal should be "self-evolution." The statement immediately elevated a conversation that included top domestic players like Moonshot AI and Zhipu AI, marking not only a bold declaration of Xiaomi's technical roadmap but also introducing a new dimension to the breakneck large model race.
Is 'AGI Is Here' a Bold or Pragmatic Claim?
With industry leaders like Moonshot AI CEO Yang Zhilin and Zhipu AI CEO Zhang Peng in attendance, Luo's assertion that AGI is a reality was particularly striking. He argued that current large models have reached or even surpassed human-level performance in specific domains, which in a sense marks the arrival of AGI. He further suggested that the industry's focus should shift from merely improving model capabilities to enabling AI to learn and iterate on its own—a concept he calls "self-evolution."
This view challenges the prevailing consensus that AGI remains a distant goal. However, Luo's other comments suggest his claim may not be baseless. He noted that the power of the open-source community, particularly through AI agent frameworks like OpenClaw, has significantly raised the performance ceiling for second-tier closed-source models. In many scenarios, their task completion rates are now very close to those of top-tier models. This implies that the capability gap is closing fast, and the key to competition is no longer a single model's power but the evolutionary capacity of the entire system.
Backed by Data: The Prowess of Xiaomi's Chart-Topping MiMo Model
Luo's confidence is not without foundation. Around the time of the forum, Xiaomi's MiMo-V2-Pro model delivered a stunning performance on the international evaluation platform OpenRouter. According to the latest weekly report, MiMo-V2-Pro became the first model in the platform's history to surpass 3 trillion tokens in weekly consumption, claiming the top spot on the leaderboard. Notably, the top six positions were all occupied by Chinese large model companies, including Zhipu AI, MiniMax, StepFun and DeepSeek.
This achievement is powerful evidence of Xiaomi's technical expertise and engineering strength in the large model field. Evolving from a smartphone manufacturer to a top contender in the global AI arena, Xiaomi's rapid rise has reshaped the competitive landscape of China's AI industry. The 3 trillion token consumption figure not only represents a massive scale of user engagement but also reflects the model's effectiveness and popularity in real-world applications, providing solid market validation for Luo's "AGI is here" thesis.
The New Battlefield: From Model Wars to AI Agent Ecosystems
A clear trend at the ZGC Forum discussion was the shift from a pure contest of model capabilities to a broader exploration of AI agent ecosystems. In addition to model developers, panelists included representatives from AI infrastructure firm WUWEN Core and the open-source agent framework Nanobot.
Zhipu AI's Zhang Peng likened agent frameworks like OpenClaw to "scaffolding," arguing that they dramatically lower the barrier to entry for using top-tier models, allowing users without programming skills to realize complex ideas through simple conversation. Moonshot AI's Yang Zhilin also showed keen interest in AI agents, using the topic to guide the entire discussion.
However, the evolution toward agents brings new challenges. Zhang Peng of Zhipu AI admitted that completing complex tasks can consume 10 to 100 times more tokens than simple Q&A, causing costs to skyrocket. Xia Lixue, co-founder of WUWEN Core, added that her company's token volume has doubled every two weeks since late January, placing immense strain on underlying infrastructure. This indicates that the AI industry's competition is expanding beyond the algorithm layer to a full-stack battle encompassing systems engineering and infrastructure.
Industry Outlook: Beyond Tokens to Application and Evolution
The dialogue at the ZGC Forum clearly sketched out a roadmap for China's AI industry: once foundational model capabilities catch up and reach a high plateau, the deciding factor will no longer be marginal percentage gains on a leaderboard. Instead, victory will belong to whoever can first build a thriving AI agent ecosystem and solve the accompanying challenges of cost and efficiency.
Luo Fuli's concept of "self-evolution" points toward an even more ultimate direction for this race. As the capabilities of models themselves begin to saturate, the key will be enabling AI systems to learn, iterate and optimize through continuous interaction with the real world, much like a living organism. From this perspective, the competition between companies like Xiaomi, Zhipu AI and Moonshot AI has transitioned from a 100-meter dash to a marathon that tests endurance, strategy and the ability to build a robust ecosystem.