Baidu has unveiled ERNIE 5.5 Flash Base, a next-generation sparse MoE pretrained foundation model, and has made its model weights publicly available. The model features 280B total parameters, with 14B activated per token, and was pretrained from scratch on 33.5T tokens. It supports a maximum training context of 128K, which can be extended to 1M tokens during fine-tuning and inference. Baidu has partnered with vLLM, SGLang, and others to provide open-source ecosystem support. Overall, the model outperforms mainstream foundation models of the same class across multiple benchmarks, including knowledge, reasoning, mathematics, code, multilingual, and long-context tasks. It also delivers multiple technological innovations in core components such as attention, MoE routing, and pretraining objectives, while incorporating numerous optimizations in residual connections, the output head, tokenizer, training infrastructure, and other areas. As a Base model with comprehensive general capabilities, it enables developers to pursue secondary development, such as domain adaptation and instruction fine-tuning. The official release of ERNIE 5.5 Flash will also be available soon.