Baidu: ERNIE 4.5 VL 424B A47B (retires Oct 8)

ERNIE 4.5 VL represents Baidu's approach to large-scale vision-language intelligence, built on a Mixture-of-Experts architecture that allows the model to scale to hundreds of billions of parameters while maintaining computational efficiency. The design philosophy centers on heterogeneous MoE structures with modality-isolated routing, enabling text and visual processing pathways to develop specialized representations without interfering with one another. This architectural choice is paired with techniques like router orthogonal loss and multimodal token-balanced loss, which ensure balanced utilization across expert pathways. The result is a model that achieves cross-modal knowledge fusion while preserving dedicated parameter spaces for individual modalities—effectively maintaining or even improving text-focused performance while dramatically expanding visual reasoning, document understanding, chart interpretation, and visual question answering capabilities.

The model's training leverages Baidu's PaddlePaddle deep learning framework, where pre-training achieves a Model FLOPs Utilization of 47%—a notable efficiency milestone for models of this scale. This efficiency reflects the heterogeneous MoE design, which enables the system to activate only the relevant expert pathways rather than engaging the full parameter set for every token. Experimental results demonstrate state-of-the-art performance across both text and multimodal benchmarks, with particularly strong results in instruction following and world knowledge tasks. As an open-weights release, ERNIE 4.5 VL offers researchers and developers direct access to experiment with and fine-tune Baidu's vision-language approach, making it practical for applications ranging from document processing pipelines to multimodal research initiatives that require robust visual understanding combined with strong textual reasoning.