ERNIE 4.5

Baidu's flagship multimodal foundation model

Free Multimodal
Visit Product Page →
128000 context tokens
0.3B-424B (10 variants, MoE) parameters
Apache 2.0 license
Jun 2025 released

ERNIE 4.5 is the model family Baidu unveiled in March 2025 and then open-sourced under an Apache 2.0 license on June 30, 2025, marking a shift from the closed approach of earlier ERNIE generations. The family spans ten variants built on a mixture-of-experts architecture, from a 0.3B dense model up to a 424B-parameter multimodal flagship (ERNIE-4.5-VL-424B-A47B) that activates around 47B parameters per token, with most variants supporting context windows up to about 128,000 tokens. Baidu released the weights on Hugging Face, GitHub, and its own PaddlePaddle platform, and reported benchmark results competitive with leading proprietary multimodal models. The open release was widely read as Baidu’s most aggressive move yet in the competition with DeepSeek and other Chinese open-weight labs, and Baidu has since moved on to ERNIE 5.0.