Ant group launches open-source multimodal model Ling-3.0-flash-VL with visual feedback loop

Ant Group has released Ling-3.0-flash-VL, the first native multimodal model in its Bai Ling series. The model, built on the MoE architecture of Ling-3.0-flash, has 124B parameters and activates 5.5B during inference. It supports image, text, and video inputs with a 256K token context window. The model introduces a visual feedback loop, enabling tasks like medical report interpretation and GUI automation through an 'observe → act → validate → correct' cycle. Training with native multimodal data improved its text capabilities, achieving a 4-point increase in the Artificial Analysis Intelligence Index v4.1.1 compared to the text-only version. Ling-3.0-flash-VL, available on Ling Studio, offers BF16 and FP8 versions on Hugging Face and ModelScope, with FP4 and INT4 versions to be released soon. The model outperformed GPT-5.4 in Image-to-WebDev Arena and excels in real-time video dialogue and long-task processing.