Baidu Open-Sources ERNIE 5.5 Flash Base With 280B Parameters

Baidu's decision to release a 280B-parameter mixture-of-experts base model openly puts a frontier-scale architecture into the hands of anyone who wants to fine-tune it, which matters most for developers and smaller labs that cannot pretrain at this scale themselves. The 14B active-parameter design keeps inference comparatively lean relative to total size, and the jump from a 128K pretraining context to 1M-token support in fine-tuning and inference signals that long-context work is where Baidu expects downstream differentiation. Whether open base weights at this scale become a recurring pattern among Chinese labs, rather than a one-off, is the open question worth watching.