Baidu Releases Open Vision-Language MoE Model

Chinese tech giant Baidu has released ERNIE 4.5 VL, a powerful new vision-language model available under a permissive open-source license. The model is designed for complex reasoning tasks that require understanding both images and text, positioning it as a capable new entry in the open multimodal space.

Efficient by Design

At its core, ERNIE 4.5 VL is a sparse Mixture-of-Experts (MoE) model. While it contains a total of 28 billion parameters, it only activates a fraction—around 3 billion—for any given inference task. This design, hinted at by the 'A3B' (Active 3 Billion) in its name, aims to provide the power of a much larger model while maintaining greater computational efficiency during use.

The model's full name, ERNIE 4.5 VL 28B A3B Thinking, emphasizes its focus on multi-step reasoning. It's built to analyze visual information and perform logical thinking, a challenging frontier for AI development.

By releasing this model under the Apache 2.0 license, Baidu is making a notable contribution to the open-source ecosystem. This gives researchers and developers a sophisticated, efficient, and freely usable foundation for building the next generation of multimodal applications.