Ant Group has added another native multimodal model to a crowded Chinese release schedule. Ling-3.2-flash landed on 5 November and takes text, images and video through one model rather than routing them to separate systems.
The specification is unusual at this size. Across a mixture-of-experts design the model holds 560 billion parameters in total, wakes roughly 25 billion of them for each inference step, and accepts a context window that stretches to one million tokens.
Where Ant thinks the model earns its keep
Four workloads get the emphasis from the company: complex visual understanding, multi-step agent work, 3D understanding and interaction, and medical imaging tasks. Those are applications where holding a long visual record in context matters more than raw benchmark scores.
Access is being handled pragmatically. A free two-week window runs after launch, and developers can reach the model either through an online chat interface or through APIs that mimic the OpenAI and Anthropic calling conventions. That compatibility layer is a deliberate bid to lower switching costs for teams already building against Western APIs.
The competitive picture
Ant is releasing into a field where Tencent shipped Hy4 with vision and computer-use capabilities this week, DeepSeek is preparing a plugin marketplace for its Harness product, and Moonshot has been publishing model weights openly.
For startups building on Chinese models, the practical question is less about parameter counts than about whether the free window turns into a retained user base. Ant has not disclosed pricing after the trial period, nor the hardware requirements for running Ling-3.2-flash, which determines whether the model is something a small team can host or only rent.
