Ant Group open-sources native multimodal model Ling-3.0-flash-VL

Ant Group has launched and open-sourced Ling-3.0-flash-VL, its first native multimodal model from the Bailing series. Based on the MoE architecture, it has 124B total parameters, activating only 5.5B per inference, and supports image, text, and video input with a 256K token context window.

The model introduces a visual feedback loop, turning tasks into a cycle of "observe, act, verify, refine," which helps in areas like reading medical reports, generating front-end code, and GUI automation. In tests, it scored 4 points higher than the text-only version on the Artificial Analysis Intelligence Index.

For image-to-webpage tasks, it understands layouts and generates code, then checks and fixes itself by comparing rendered results. Its score in the Image-to-WebDev Arena was higher than GPT-5.4. As a GUI agent, it can recognize interface structures and operate across tools. In medical reports, it pulls data from multiple documents and flags risks. With an any-resolution visual encoder and VideoRoPE, it responds quickly in real-time video chats and long tasks.

Ling-3.0-flash-VL is now free to try on Ling Studio, and the BF16 and FP8 versions are open on Hugging Face and ModelScope. FP4 and INT4 versions are coming soon.