Baidu Releases ERNIE 5.5 Flash Preview and Plans Open-Weight Release
Baidu released ERNIE 5.5 Flash Preview, the first post-trained version of ERNIE 5.5 Flash Base, and made its API available through the Qianfan platform. Baidu says the sparse MoE model has 280B total parameters, 14B active parameters, a 1M-token context window and generation speed above 200 tokens per second. The company also reports competitive results in coding, 3D modeling, knowledge work and content creation. Baidu says the model weights will be released to the global open-source community by November 25 under the Apache 2.0 license. It warns that the Preview remains an intermediate post-training version that may still have capability defects, with further training and optimization planned.
The release makes an efficiency-focused Baidu model available for testing through Qianfan while setting a planned deadline for publishing its weights openly. Baidu's specifications, speed and benchmark results have not been independently verified in the supplied reports.