Baidu open-sources its ERNIE 5.5 Flash Base model

Baidu published the weights for ERNIE 5.5 Flash Base on Hugging Face and ModelScope, making a large Chinese-developed language model freely downloadable. The model uses a sparse mixture-of-experts design with about 280 billion total parameters but only around 14 billion active per token, which keeps inference costs manageable despite its size. It was pre-trained on roughly 33.5 trillion tokens and supports a 128,000 token context window that can be expanded toward 1 million tokens for longer documents. Baidu released both FP8 and BF16 versions of the checkpoints, and early support has already appeared for popular inference frameworks vLLM and SGLang, which should make it easier for developers to run the model efficiently on their own hardware. The release adds to a growing list of large open-weight models coming out of Chinese labs this year, continuing a trend of major players treating open releases as a way to build developer mindshare globally.

Why it matters

Another major Chinese lab opens up a large, efficient model, expanding options for developers who need open weights.

Source: Baidu