Ant Ling launches a multimodal model for million-token visual agents

Ant Group’s Ant Ling team has released Ling-3.2-flash, a native multimodal mixture-of-experts model built for visual understanding and long-running agents. The model has 560 billion total parameters but activates about 25 billion for each token, a design intended to combine large capacity with lower inference cost.

Ling-3.2-flash accepts text, images and video and supports a context window of up to one million tokens. Ant Ling positions it for complex document analysis, interface automation, 3D environments and medical images. Its developer documentation emphasizes agents that can observe visual state, plan across many steps and act inside software rather than only describe an image.

The company reports that the model beats Claude Opus 5 on selected document, visual reasoning and image-to-web benchmarks. Those numbers are vendor-run release evaluations, so independent testing will be needed before treating them as a reliable ranking. Ant Ling’s launch materials also show demos involving frontend generation, Blender operations and document-to-code workflows.

The model is available through Ant Ling’s API. Public model weights were not available at launch, despite earlier Ling releases being described as open-source. For builders, the practical change is access to a large-context visual agent model with relatively sparse activation; the open-weight status and real-world reliability remain open questions.