Ling-3.0-flash-VL

Ling-3.0-flash-VL is a multimodal large language model (MLLM) developed by inclusionAI, the AI research laboratory of Ant Group. Released in September 2026, it is the first vision-integrated model within the Ling-3.0-flash family. The model is designed to natively process and reason across text, image, and video inputs, supporting applications such as document intelligence, medical report interpretation, and agentic workflows that require visual feedback for GUI navigation.

The model utilizes a sparse Mixture-of-Experts (MoE) architecture with a total of 124 billion parameters, of which 5.5 billion are activated per token during inference. Its 42-layer hybrid backbone, identified as BailingMoeV3VL, alternates between Knowledge-Dense Attention (KDA) and Gated Multi-head Latent Attention (MLA) layers at a 5:1 ratio. This architectural choice is optimized for inference efficiency and long-context stability, enabling the model to support a context window of up to 1 million tokens.

Visual processing is powered by a ViT visual encoder and a two-layer MLP projector that aligns visual features with the model's textual representations. To handle the temporal dimension of video data, Ling-3.0-flash-VL employs VideoRoPE (Rotary Positional Embedding), which allows the model to encode both spatial positions and the temporal order of frames. This facilitates advanced capabilities such as event localization, temporal reasoning, and visual change detection over time. The weights for Ling-3.0-flash-VL are released under the MIT license, supporting both open-source research and commercial integration.