Baidu open-sources the ERNIE 5.5 Flash model with 280 billion parameters.

ChainCatcher report: Baidu has open-sourced the model weights of Wenxin ERNIE-5.5-Flash-Base. Following the Wenxin 4.5 series in June 2025, Baidu has once again open-sourced a trillion-parameter large model after approximately 16 months; during this period, the released Wenxin 5.0 and 5.1 models were not open-sourced. Wenxin 5.5 Flash is optimized for inference efficiency, with a total of 280 billion parameters, activating only 14 billion parameters per generated token, and pre-trained from scratch on 33.5 trillion tokens. The model supports an inference context length of up to 1 million tokens. Baidu has redesigned its long-text processing mechanism: among the 43 network layers, 37 layers compress every 128 historical tokens into a single cached entry, while the remaining 6 layers retain fine-grained retrieval capabilities over the original content to reduce memory and computational overhead for long-text processing. In unified benchmark evaluations against DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, it ranked first in 24 out of 30 benchmarks. This release includes only the pure text-based Base model without fine-tuning or post-training, primarily intended for developers to perform domain-specific fine-tuning, instruction tuning, and reinforcement learning. Baidu stated that the post-trained version of Wenxin 5.5 Flash will be released subsequently. The license is stricter than its predecessor: while Wenxin 4.5 used the Apache 2.0 license, Wenxin 5.5 Flash uses a custom license that prohibits use in the European Union, the United Kingdom, and South Korea; commercial use by entities with revenues exceeding $50 million in the previous 12 consecutive months requires separate written authorization from Baidu.