Ant Group releases Ling-3.2-flash: 560B-parameter multimodal model with 1M-token context

Ant Group released Ling-3.2-flash on November 5, a native multimodal model that accepts text, images, and video. The model has 560 billion total parameters, activates 25 billion per inference step, and supports a one-million-token context window, according to TechNode's report on the launch.

Key specifications

• Model: Ling-3.2-flash
• Developer: Ant Group
• Release date: November 5, 2026
• Input modalities: text, images, video
• Total parameters: 560 billion
• Active parameters: 25 billion per inference step
• Context window: one million tokens

Stated focus areas

Ant says the model focuses on complex visual understanding, multistep agent tasks, and 3D understanding and interaction. These are Ant's own descriptions of the model's emphasis. No benchmark scores accompany them in the report.

What this means

Ling-3.2-flash combines a one-million-token context window with native input across text, images, and video. The gap between 560 billion total parameters and 25 billion active per inference step means only a fraction of the model's weights are used for each step. Ant's stated emphasis on multistep agent tasks and visual understanding indicates the model is aimed at workloads that combine long inputs with several modalities. Independent evaluation will be needed to test those claims.

This article is based on TechNode's report of the launch.