Baidu released ERNIE 5.5 Flash Base, a 280B-parameter hybrid-attention MoE model with 14B active parameters. Pretrained on 33.5T tokens, it is positioned as an open and efficient pretraining base.