Ant Ling releases Ling-3.2-flash, a 560B multimodal MoE

Ant Ling’s account introduced Ling-3.2-flash on November 5, 2026, calling it a native multimodal model with 560 billion total parameters and 25 billion active. The Ling developer docs describe the same release: a Mixture-of-Experts model with the context window capped at 1 million tokens and native support for image, text, and video.

The post and the docs name the same focus. Ant Ling said the model is stronger on complex visual understanding, long-horizon multimodal agents, 3D understanding and interaction, and medical intelligence. The docs call the fourth area medical multimodal, and say the point is to join multimodal perception, deep reasoning, and autonomous execution on complex real-world tasks.

Documentation is explicit that this is not a captioning add-on. Image, text, and video are foundational inputs for reasoning, and vision and language are fused in the architecture rather than passed to an external vision module. Stated uses include visual reasoning over images and video, multi-step agent work that stays aligned with a goal, understanding 3D scenes and spatial relationships, and medical imaging understanding with report interpretation.

On compute, the docs contrast Ling’s MoE with dense models that activate every parameter. Each token is routed to the most relevant expert subnetworks, so capacity and per-step compute are not the same number. Ling-3.2-flash is listed at 560B total and about 25B activated per step. The 1M window is described as support for long documents and long video. A series-level claim says long-range retrieval does not noticeably degrade whether the needed information is at the beginning, middle, or end of the context.

In the selection table, that 560B / 25B / 1M shape is shared with Ling-3.1-flash, which the docs position for office work, long-horizon software development, and complex agent tasks rather than as the native multimodal model. The previously released native multimodal model, Ling-3.0-flash-VL, is listed at 124B total and 5.5B activated, with a 256K window. The evolution table dates the Ling 3.2 series to November 2026.

The release post says Ling-3.2-flash outscores Opus 5 (max) in release evaluations. That is Ant Ling’s claim, not an independent ranking. The wording of the post does not include the underlying scores, and the model docs do not publish a comparison against Opus 5.

For the spec in one place, read the Ling model page, including the documented path to an API key and a first call, then the November 5 account post for the release wording.

Sources