BiTui News: Baidu has open-sourced the model weights of ERNIE-5.5-Flash-Base. Following the release of the ERNIE 4.5 series in June 2025, Baidu has once again open-sourced a trillion-parameter large model after approximately 16 months. The subsequent ERNIE 5.0 and 5.1 models were not open-sourced.
Baidu's Wenxin 5.5 Flash emphasizes reasoning efficiency, with a total of 280 billion parameters, activating only 14 billion parameters per generated token. The model was pre-trained from scratch on 33.5 trillion tokens and supports up to 1 million tokens in inference context.
Baidu has also redesigned its long-text processing mechanism. In a 43-layer network, 37 layers compress every 128 historical tokens into a single cached entry, while the remaining 6 layers retain fine-grained access to the original content, reducing memory and computational overhead for long-text processing.
It was uniformly evaluated alongside DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, ranking first in 24 out of 30 benchmark tests.
This release is the pure text-based Base model without pre-training or post-training, primarily intended for developers to continue fine-tuning for specific domains, instruction training, and reinforcement learning. Baidu stated that the post-trained ERNIE 5.5 Flash will be released shortly.
The license is stricter than the previous generation. ERNIE 4.5 uses the Apache 2.0 license, while ERNIE 5.5 Flash has switched to a custom license that does not permit use in the European Union, the United Kingdom, and South Korea. Products or services with revenues exceeding $50 million over the prior 12 consecutive months must obtain separate written authorization from Baidu before commercial use.