Baidu open-sourced Wenxin 5.5 Flash, with a total parameter of 280 billion

Baidu has open-sourced the ERNIE-5.5-Flash-Base model weights. Following the release of the ERNIE 4.5 series in June 2025, Baidu has once again open-sourced a large model with hundreds of billions of parameters after approximately 16 months, during which the ERNIE 5.0 and 5.1 versions were not open-sourced. The ERNIE 5.5 Flash focuses on inference efficiency, with a total of 280 billion parameters, activating only 14 billion parameters for each generated token, and using 33.5 trillion tokens for zero pre-training.

This model supports a maximum inference context of 1 million tokens. Baidu has redesigned the long text processing mechanism, with 37 out of 43 layers compressing every 128 historical tokens into a single cache record, while the remaining 6 layers retain fine retrieval capabilities for the original content, reducing the memory and computational overhead of long text processing. The official evaluation was conducted in conjunction with DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base, and Xiaomi MiMo-V2.5-Base, ranking first in 24 out of 30 benchmark tests.

What is being opened this time is a pure text Base model that has not undergone intermediate training or post-training, mainly for developers to continue domain fine-tuning, instruction training, and reinforcement learning. Baidu stated that the post-trained ERNIE 5.5 Flash will be released subsequently. The license is stricter than the previous generation; the ERNIE 4.5 used the Apache 2.0 protocol, while 5.5 Flash has switched to a custom license that does not authorize use in the European Union, the United Kingdom, and South Korea; for related products or services that have generated over 50 million dollars in revenue in the past 12 consecutive months, separate written authorization from Baidu must be obtained before commercial use.