Ant Group has officially announced the launch and open-sourcing of Ling-3.0-flash-VL, the inaugural native multimodal large model within its Bailing series. Leveraging the Mixture of Experts (MoE) architecture of Ling-3.0-flash, this model boosts its total parameters to an impressive 124 billion, while only activating 5.5 billion parameters during a single inference, optimizing computational efficiency. It seamlessly supports multimodal inputs, including images, text, and videos, and features a context window capable of processing up to 256K tokens, facilitating comprehensive understanding and generation.

A standout feature of Ling-3.0-flash-VL is its integration of a visual feedback closed-loop mechanism. This innovative approach enables the model to dynamically adjust and correct its execution results based on visual feedback across diverse scenarios, enhancing accuracy and adaptability. Through native multimodal joint training, the model has significantly elevated its textual processing capabilities, surpassing both its text-only counterpart and GPT-5.4 in relevant benchmarks, demonstrating superior performance in complex language tasks.

Ling-3.0-flash-VL excels in executing intricate tasks, such as transforming images into fully functional web pages, performing cross-tool GUI operations with precision, and interpreting medical reports with high accuracy. These capabilities are underpinned by its adoption of an arbitrary-resolution visual encoder and VideoRoPE technology, coupled with a 42-layer hybrid architecture serving as its linguistic backbone. This combination ensures swift, low-latency responses, making it ideal for real-time applications.

Currently, users can explore Ling-3.0-flash-VL for free on Ling Studio. The model is available in BF16 and FP8 versions on relevant open-source platforms, with plans to release FP4 and INT4 versions in the near future, further expanding its accessibility and versatility.