Baidu releases ERNIE 5.5 Flash Base, described as a raw pretraining checkpoint for further training. The model is presented as a mixture-of-experts (MoE) system with 280B total parameters and 14B active parameters during use.
Both outlets say the release comes with a 128K training context and support for up to 1M tokens. MarkTechPost additionally frames the checkpoint as being published before any mid- or post-training steps, and notes it was trained on 33.5T tokens. Pandaily focuses on the availability of the pretraining checkpoint and its configuration, without adding the training-data figure.
Overall, the coverage aligns on the model’s size, MoE active-parameter count, and long-context/token-support specifications, with differences mainly in the amount of background MarkTechPost provides about prior training and release timing.
