Ling 3.0 Flash VL is a native multimodal model released by inclusionAI, the AI organization associated with Ant Group, and distributed openly through the inclusionAI spaces on Hugging Face and ModelScope. It is positioned as the next-generation vision-language follow-on to Ling 3.0 Flash, inheriting that predecessor's language, reasoning, and long-context strengths while adding native image and video understanding. The model is aimed at workflows where visual information needs to feed directly into reasoning and acting, rather than being treated as a separate perception stage.

Under the hood, Ling 3.0 Flash VL combines a ViT visual encoder with a two-layer MLP projector that aligns image and video features with the language representation space, and it uses VideoRoPE to encode both spatial positions and temporal order so it can reason over visual change across frames. Its 42-layer hybrid backbone alternates KDA and Gated MLA layers at a 5:1 ratio, supporting long multimodal contexts while keeping compute efficient. With 124B total parameters and only about 5.5B activated per token, it targets practical deployments that need strong multimodal reasoning, video question answering, event localization, and agentic tool use over very long inputs, all in an open-weights package.