By 2026, the wave of artificial intelligence is shifting from simple conversational interactions toward a more disruptive era of intelligent agents. In this transition of technological paradigms, Xiaomi, led by a so-called "AI prodigy," is making a strong entrance with its trillion-parameter MiMo large model and has put forward an ambitious prediction—the most critical trend in the AGI process over the next year will be "self-evolution." This statement comes from Luo Fuli, head of Xiaomi's MiMo large model.
Luo, a former DeepSeek employee known for her youth and sharp technical insights, is leading Xiaomi's core team—with an average age of just 25 and over 60% graduates from Tsinghua and Peking University—to explore a new path in AI's uncharted territory, distinct from the industry's focus on "stacking parameters and competing on computing power." Backing her efforts is Lei Jun's bold bet of over 60 billion yuan on AI in the next three years, reflecting Xiaomi's ultimate ambition to transform from a hardware company into a global hardcore technology firm.
Luo Fuli's "Self-Evolution" Declaration "I personally view this type of agent framework as a very revolutionary and disruptive event. If I had to choose one word to summarize the most critical development in the AGI process over the next year, it would be 'self-evolution,'" Luo stated in her latest public remarks. In her view, "self-evolution" is no longer a sci-fi concept but a reality already taking shape in the lab. She revealed that top-tier models can now autonomously optimize for specific tasks and run stably for two to three days. Researchers on her team, using advanced models like Claude Code, have seen "research efficiency improve nearly tenfold." This explosive gain in efficiency convinces her that self-evolution is the true direction for creating entirely new things. "It’s not simply about replacing existing human productivity but, like top scientists, exploring things that don’t yet exist in the world," Luo explained.
She believes that previous dialogue-based paradigms fall far short of realizing the full potential of pre-trained models, and that agent frameworks are now activating this upper limit. A robust self-evolution agent framework could bring "exponential acceleration" to scientific research. More radically, Luo has revised her timeline predictions: "A year ago, I thought this process would take three to five years; now, I believe the window has shortened to one to two years." In her view, AI's evolution has far exceeded expectations, and "we are already touching this boundary." Such a confident, aggressive outlook contrasts sharply with the industry's generally cautious optimism—raising the question of whether AI can truly achieve scientist-level autonomous exploration within a year or two. This is not just a technical debate but a fundamental divergence in understanding the nature of AGI.
Lei Jun's 60 Billion Yuan Bet In contrast to Luo Fuli's pioneering "youthful ambition," Lei Jun's approach to Xiaomi's AI strategy has been characterized by a low-profile, results-first mentality. Unlike peers such as Tencent and Alibaba, which invested early in large models, Xiaomi waited until March 2026 to reveal its hand all at once: launching the trillion-parameter MiMo-V2-Pro MoE model, ranked eighth globally and fifth in brand rankings; simultaneously introducing the full-modal V2-Omni and voice model V2-TTS to build a complete agent technology stack. Lei Jun announced that "Xiaomi will invest over 60 billion yuan in AI over the next three years," with R&D and capital expenditure exceeding 16 billion yuan in 2026 alone. Rough comparisons show that AI investment already exceeds half of Xiaomi's spending on car manufacturing, underscoring AI's central role in the company's strategy.
Lei's logic is clear: AI represents the incremental opportunity of the era, a core technology Xiaomi must master for the next decade. However, a subtle tension exists between Lei's pragmatism and Luo's idealism. Lei emphasizes technology implementation and ecosystem integration. The primary mission of the MiMo model is to deeply empower Xiaomi's "human-car-home full ecosystem": from AI-native products like "miclaw" on smartphones, to intelligent driving in cars, and seamless coordination in smart homes. His focus is on immediately translating AI into product experience and market competitiveness. Luo, meanwhile, appears more concerned with the model's underlying capabilities and scientific essence. She has publicly questioned the industry's "brute force" approach, arguing that computing power and data are not the ultimate moats—instead, scientific research culture and methodology are. Her goal is to develop models that understand the physical world, possess spatiotemporal coherence, and ultimately create an "AI virtual universe" on the path to AGI.
Nevertheless, the AI exploration led by Luo Fuli's team is commendable. Under her guidance, the MiMo team has chosen a differentiated path: avoiding blind parameter escalation and instead focusing intensely on agent and reasoning efficiency. Although MiMo-V2-Pro has a trillion total parameters, only 42 billion are activated, using an innovative hybrid attention architecture optimized for high-intensity agent scenarios. It supports million-level context windows and achieves an 81% task completion rate, rivaling top models. At the same time, Xiaomi is actively open-sourcing; the MiMo-7B series is already available, and V2-Flash, with its low cost and high speed (three times faster inference than DeepSeek), ranks second among global open-source models.
Reflecting on Luo Fuli's response to Yang Zhilin's question—"If I had to choose one word to summarize the most critical development in the AGI process over the next year, it would be 'self-evolution'"—the industry can look forward to witnessing the results of Xiaomi and Luo's "self-evolution" endeavors in 2026.
Comments