ERNIE 5.1 is the latest model in the Wenxin series, with comprehensive upgrades to its foundational capabilities and significant improvements in agents, knowledge, reasoning, and deep search. This upgrade uses a decoupled fully-asynchronous reinforcement learning technique to specifically address challenges encountered as large models evolve toward agent-based autonomous decision-making, such as training–inference numerical bias, low utilization of heterogeneous resources, and global issues caused by long-tail effects. It is paired with scaled agent post-training techniques to enhance model capabilities and generalization, enabling a three-step collaboration of environment, expert, and fusion that both ensures training efficiency and significantly improves the model’s stability and performance on complex tasks.