ERNIE 4.5 Full Family Open-Sourced: Baidu Uses Heterogeneous MoE to Complete Multimodal Open Route
Baidu open-sourced 10 model variants of ERNIE 4.5 in June 2025 under the Apache 2.0 license.
Wenxin large model has shifted from primarily relying on its own products and cloud services to simultaneously offering a complete open model family. Baidu is now leveraging the PaddlePaddle ecosystem and multi-hardware deployment to attract developers.
ERNIE 4.5 includes text and vision-language models, with a maximum of 424B total parameters and 47B activated parameters, using a heterogeneous MoE that combines cross-modal shared and dedicated parameters.
Major companies are beginning to offer both closed-source APIs and open weights. The complete model series, training framework, and hardware compatibility collectively influence competition in the Chinese developer ecosystem.
Enterprises can compare between Baidu Cloud and self-deployment, but should unify task sets, hardware costs, and version upgrade metrics to avoid focusing solely on model parameters.
Observe open-source repository maintenance, inference framework compatibility beyond Paddle, overseas adoption, and capability differences between commercial APIs and open versions.