ERNIE-4.5-VL-28B-A3B-Thinking

This model utilizes a Mixture-of-Experts architecture, combining a 21-billion parameter language decoder with a 7-billion parameter vision encoder to process complex visual and textual data. By activating only 3 billion parameters per token, it achieves a balance between high-performance reasoning and computational efficiency. The design is specifically intended to handle nuanced visual-textual tasks, such as extracting on-screen text, analyzing video content over time, and executing tool-based actions like image zooming or searching to resolve fine-grained details.

The model underwent an extensive mid-training phase, absorbing a diverse corpus of visual-language reasoning data to deepen semantic alignment between modalities. It incorporates multimodal reinforcement learning, utilizing GSPO and IcePop strategies to stabilize its routing architecture alongside dynamic difficulty sampling for improved learning efficiency. These post-training advancements enable the model to excel in STEM problem-solving and visual grounding, providing a robust, Apache 2.0-licensed foundation for developers seeking to integrate sophisticated multimodal analysis into commercial applications.