Luo Fuli, the distinguished head of Xiaomi’s MiMo large model team, has articulated a fundamental shift in the artificial intelligence paradigm, moving from a "Chat era" dominated by pre-training to an "Agent era" where post-training takes precedence. This insight emerged from an extensive interview, where Fuli, a former core author of DeepSeek V2 and an alumna of Alibaba DAMO Academy, shared her observations on the evolving AI landscape. Her move to Xiaomi in November 2025 to lead its MiMo team was a high-profile recruitment, underscoring the fierce competition for top AI talent.
A key indicator of this paradigm shift, according to Fuli, is a dramatic reallocation of computational resources. She revealed that the compute allocation for research, pre-training, and post-training, which was approximately 3:5:1 in the Chat era, has now shifted to a more balanced 3:1:1. This signifies that the investment in post-training, crucial for developing sophisticated AI agents, has reached parity with pre-training in leading teams.
Fuli underscored the transformative impact of agent frameworks, citing her personal experience with OpenClaw, an open-source AI agent that gained significant traction in early 2026 for its ability to execute complex tasks autonomously. She described being "excited to the point of not sleeping" after using OpenClaw, noting its "soulfulness" and ability to perform tasks previously thought impossible. Her team at MiMo, under her directive, rapidly adopted OpenClaw, leading to an explosion of collective intelligence that accelerated research output significantly.
The MiMo team's focus on agent scenarios is deliberate, with their MiMo-V2.5 models specifically designed for intelligent agent tasks, offering full-modal capabilities. These models have demonstrated strong performance in agent benchmarks, with the MiMo-V2-Pro even being mistaken for DeepSeek V4 earlier in 2026, highlighting its advanced capabilities. Fuli stated that a well-orchestrated agent framework can substantially compensate for model shortcomings, enabling even mid-tier models to achieve performance comparable to advanced systems like Claude Sonnet.
Regarding the competitive landscape, Fuli asserted that Anthropic's approach is widely recognized as correct, and Chinese teams have largely closed the gap in pre-training. She identified 1-trillion-parameter base models as the "entry ticket" to achieving performance akin to Claude Opus 4.6, with the next critical battleground being the scaling of reinforcement learning within agent scenarios. Fuli also shared an aggressive outlook on Artificial General Intelligence (AGI), predicting it could reach 60-70% realization by the end of the current year and full achievement within two years.
