Ant Group open-sources the multimodal model Ling-3.0-flash-VL with 124 billion parameters.

Huoxing Finance reports that on September 9, Ant Group announced the official release and open-sourcing of Ling-3.0-flash-VL, the first native multimodal large model in the Ling series. Built upon the MoE architecture of Ling-3.0-flash, the model has a total of 124B parameters, activating 5.5B parameters per inference, natively supporting image, text, and video inputs, with a context window of 256K tokens. Focusing on “how to complete real-world tasks more reliably and efficiently,” Ling-3.0-flash-VL explores the following directions: A common concern when adding visual capabilities to large models is that they may degrade textual intelligence. However, training practices have yielded the opposite conclusion: native multimodal joint training not only expands the application boundary but also enhances textual intelligence. Ling-3.0-flash-VL introduces a visual feedback mechanism—observing execution outcomes, comparing against goals, identifying deviations, and continuously correcting—that transforms tasks from one-time “generation” into a closed loop of “observe → act → verify → correct,” resulting in more reliable outcomes. Ling-3.0-flash-VL inherits the core advantage of Ling-3.0-flash as an efficient execution node in agent workflows, balancing output quality and execution efficiency within the visual feedback loop to advance complete tasks at lower cost and in less time.