Baidu unveils ERNIE 5.0, a 2.4-trillion-parameter model for text, images, audio and video, at Baidu World 2025

Baidu/ERNIE 5

At its Baidu World conference in Beijing, Baidu showed ERNIE 5.0, a natively omni-modal model with 2.4 trillion parameters that both reads and produces text, images, audio and video. It was made available on Baidu's Qianfan cloud platform, and Baidu also announced new Kunlun chips for running models of this size.

• 2.4 trillion parameters; one model for understanding and generating four kinds of media.
• Offered from launch through Baidu's Qianfan cloud platform.
• Kunlun M100 inference chip due early 2026, M300 for training and inference early 2027.