10B beats 200B! Step3-VL-10B, the state-of-the-art (SOTA) open-source visual language for leaping stars.

StepFun has released its open-source 10B parameter multimodal model, Step3-VL-10B, which outperforms mainstream large-scale models with more than 20 times the number of parameters in multiple benchmark tests. The model achieves state-of-the-art (SOTA) performance in core dimensions such as visual perception, mathematical competitions, and logical reasoning. Its innovative Parallel Coordinated Inference (PaCoRe) mechanism significantly improves the ability to handle complex tasks.