According to CoinMeta, Ant Group announced on September 9th the official release and open-sourcing of the first native multimodal large model in its Bailing series, Ling-3.0-flash-VL. This model is developed based on the MOE architecture of Ling-3.0-flash, with a total of 12.4 billion parameters. It activates 550 million parameters per inference and natively supports image, text, and video inputs, with a context window of up to 256k token. The main focus of Ling-3.0-flash-VL is to explore how to complete real tasks more reliably and efficiently, especially by introducing a visual feedback mechanism that transforms the task from a one-time generation into a closed loop of observation, action, verification, and correction, thereby improving the reliability of the execution results.