ERNIE 4.5 VL 424B A47B

ERNIE 4.5 VL 424B A47B is Baidu's large multimodal Mixture-of-Experts model from the ERNIE 4.5 family, built around a heterogeneous MoE structure that jointly trains on text and visual modalities. The architecture routes modalities through isolated paths and uses router orthogonal loss along with a multimodal token-balanced loss so that image and language learning reinforce each other rather than compete. This design lets the model handle cross-modal reasoning tasks where understanding visual content in the context of language is essential, rather than treating vision as a bolted-on encoder.

With 424 billion total parameters and 47 billion activated per token, the model concentrates substantial capacity into each forward pass while keeping per-token compute manageable, which suits enterprise workloads that need strong reasoning without unlimited inference budgets. Baidu has also published a Base variant with PaddlePaddle-format weights on its public model hub, indicating an open base release path that researchers can build on for downstream vision-language experimentation. The practical fit is teams that need a flagship-class multimodal model for tasks such as document understanding, visual question answering, and grounded analysis, especially where Baidu's ecosystem of tools and deployment options is already in play.