After a 16-month hiatus, Baidu once again unveils its large model arsenal: the 280-billion-parameter Wenxin 5.5 Flash goes open source.

Beating AI News Flash: Baidu open-sources the ERNIE-5.5-Flash-Base model weights. Following the ERNIE 4.5 series in June 2025, Baidu has open-sourced a hundred-billion-parameter-class large model again after about 16 months. The ERNIE 5.0 and 5.1 released in between were not open-sourced.

ERNIE 5.5 Flash focuses on inference efficiency, with 280 billion total parameters and only 14 billion parameters activated per token generated. The model was pre-trained from scratch on 33.5 trillion tokens and supports an inference context of up to 1 million tokens.

Baidu also redesigned the long-text processing mechanism. Of the 43 layers in the network, 37 layers compress every 128 historical tokens into a single cache record, while the other 6 layers retain fine-grained retrieval capability over the original content, in order to reduce the memory and compute overhead of long-text processing.

Officially, it was evaluated uniformly against DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, ranking first in 24 of 30 benchmark tests.

What is being released this time is a pure-text Base model that has not undergone mid-training or post-training, mainly for developers to continue domain fine-tuning, instruction training, and reinforcement learning. Baidu said that the post-trained ERNIE 5.5 Flash will be launched subsequently.

The license is stricter than the previous generation. ERNIE 4.5 used the Apache 2.0 license, while 5.5 Flash has switched to a custom license and is not authorized for use in the EU, the UK, and South Korea. For related products or services with revenue exceeding $50 million over the past 12 consecutive months, separate written authorization from Baidu must be obtained before commercial use.