Xiaomi's large model team lead Luo Fuli gave a rare 3.5-hour in-depth interview on bilibili, laying out her blueprint for AI's next phase. Born in 1995, Luo has worked at Alibaba DAMO Academy, Fantuo Quant, and DeepSeek before joining Xiaomi to lead its core LLM research. Known as a legendary figure in the industry, she rarely speaks publicly. The interview packed dense insights that many AI practitioners praised as 'every minute is pure substance.'
From Chat to Agent: The Battlefield Has Shifted
Luo stated outright that the competition track for large models has changed. In the past, the Pre-train-dominated Chat era saw rivals competing on foundational capabilities and general dialogue quality. But by 2026, the battlefield has fully moved to the Post-train-dominated Agent era. Whoever enables models to autonomously execute complex tasks without human intervention will seize the next wave of market share. She emphasized that 'self-evolution' will be the most critical event for AGI in the coming year. Top models can already self-optimize on specific tasks and run stably for 2 to 3 days without human adjustment. This means AI systems are gaining a degree of self-correction ability, with boundaries quietly shifting.
Compute Ratio Overhaul: Post-Training Slimmed, Inference Elevated
Luo disclosed a major shift in Xiaomi's internal compute allocation. The industry standard had been Pre-train : Post-train : Inference = 3:5:1, with post-training consuming the most resources. Xiaomi has now adjusted to 3:1:1, sharply cutting post-training's share while boosting Pre-train and inference. The logic: Agent RL Scaling strategies are maturing, so post-training no longer requires brute-force compute. The model's self-improvement efficiency has greatly improved. The rise in inference resources reflects the high demand for real-time responsiveness in deployed Agents.
Pre-Train Gap Shrinks from 3 Years to Months, What's Next
Addressing the long-criticized 'pre-training gap' for Chinese AI teams, Luo gave a relatively optimistic view: the gap has narrowed from 3 years to months. This isn't self-congratulation but a strategic turning point. With the pre-training gap nearly closed, resource focus should move to Agent RL Scaling, which is exactly Xiaomi's current bet. She also discussed Anthropic's path and the ripple effects of 2026's Claude Opus 4.6 and OpenClaw, arguing these variables are accelerating the industry's shift from 'tool' thinking to 'agent' thinking.
MiMo-V2: Xiaomi's First Report Card
On March 19, 2026, Luo's team released the MiMo-V2 series with three models. The flagship MiMo-V2-Pro uses a trillion-parameter hybrid attention architecture with only 42B activated parameters, supports million-token context, and achieves 81% task completion. MiMo-V2-Omni targets full-modal Agent scenarios, and MiMo-V2-TTS focuses on voice synthesis. The open-source MiMo-V2-Flash grabbed second place on global open-source model leaderboards, delivering inference speed 3x faster than DeepSeek-V3.2. For a team lead who joined Xiaomi only in November 2025, this is far from a quiet debut.
Luo's career is a microcosm of AI's fast-forward history: foundational research at Alibaba DAMO, engineering efficiency at Fantuo Quant and DeepSeek (she was a key developer of DeepSeek-V2), and now complex commercial deployment at Xiaomi. This 3.5-hour interview marks her first public systematic exposition as a technical leader. Her conclusion is neither pessimistic nor glib: competition in the Agent era is far more complex than in the Chat era, and breakthroughs in 'self-evolution' are genuine technical thresholds, not marketing hype. The speed at which Chinese teams have closed gaps is encouraging, but the next battle—making agents stable and autonomous in real-world complex tasks—has only just begun.