Baidu releases ERNIE 5.1 with a cheaper agent-training stack

Baidu's ERNIE 5.1 is a good example of the non-U.S. model race shifting from raw scale to training efficiency. In its official May 9 release, Baidu says ERNIE 5.1 inherits ERNIE 5.0's pretraining foundation while compressing total parameters to roughly one-third and active parameters to roughly one-half. Baidu also claims it reaches leading performance at its model scale using about 6% of the pretraining cost of comparable models. The more interesting technical detail is infrastructure: Baidu says it built a disaggregated, fully asynchronous reinforcement-learning system to handle training-inference divergence, low resource utilization, and long-tail tasks, then used scaled agentic post-training. The model is available through the ERNIE site rather than as an open-weight release. Watch whether efficiency claims translate into cheaper inference and better agents in Baidu's search, cloud, and app ecosystem.

Key details: Baidu, ERNIE 5.1, May 9, 2026, one-third total parameters, one-half active parameters, about 6% pretraining cost, asynchronous reinforcement learning, agentic post-training.

Continue swiping for more AI Brief stories.