Luofu Li: Large models enter the post-training era, with top teams' pre-training and post-training computing power ratio reaching 1:1

According to Beating Monitoring, Luo Fuli, head of Xiaomi’s large model team, pointed out that the competition for large models has shifted from the Chat era dominated by pre-training to the Post-train dominated Agent era. The current key focus is “how to effectively scale reinforcement learning (RL) on Agents.”

This paradigm shift directly leads to a reconstruction of computing resource allocation. Luo Fuli revealed that during the Chat era, the ratio of computing power used for research, pre-training, and post-training was approximately 3:5:1; whereas in the current Agent era, a reasonable allocation ratio has become 3:1:1, meaning the investment in pre-training and post-training has become roughly equal, with top-tier model teams now investing equally in both.

At the same time, the requirements for system architecture have also undergone a major change. Previously, RL infrastructure mainly centered around the “model inference engine” to handle pure text computations; now, the infrastructure must be centered around the “Agent,” supporting heterogeneous cluster scheduling and tolerating the ambiguity caused by various uncontrollable factors that may interrupt Agents during complex workflows.