ERNIE 5.0

model paper
Baidu
Most advanced ERNIE model to date. 2.4 trillion total parameters with ultra-sparse MoE (<3% activation ratio) for extreme inference efficiency. Native full-modality unified modeling (text, image, audio, video) in a single autoregressive framework.
AnnouncementBlog PostPaper (arXiv)Artificial Analysis

Outputs 2

ERNIE 5.0

model

Most advanced model to date. 2.4 trillion total parameters with ultra-sparse MoE (<3% activation ratio) for extreme inference efficiency. Native full-modality unified modeling (text, image, audio, video) in a single autoregressive framework.

AnnouncementBlog Post
Architecture MOE
Parameters 2.4T

Previewed at Baidu World 2025 on November 13, 2025. Official release January 22, 2026.

ERNIE 5.0 Technical Report

paper

Trillion-parameter unified autoregressive multimodal model with elastic MoE training. 437 co-authors.

Paper (arXiv)Blog Post

arXiv: 2602.04705

moefrontiermultimodal