Luo Fuli, head of Xiaomi's large model team, announced a significant shift in the large model landscape from the Chat era to the Agent era, emphasizing post-training. This transition has altered compute allocation strategies, with the pre-training to post-training compute ratio now reaching 1:1 among leading teams. Previously, the ratio was 3:5:1 during the Chat era. Luo noted that the focus is now on scaling reinforcement learning for Agents, necessitating changes in system architecture to support complex workflows and heterogeneous cluster scheduling.