Ant Group's Ling-3.2-flash upgrades native vision: 560B flagship can view images to operate Blender and create 3D

Beating AI News Flash: Ant Ling releases Ling-3.2-flash, bringing native multimodal capabilities to its 560B flagship model. The new model supports image, text, and video inputs, with a focus on enhancing complex visual understanding, 3D interaction, and multimodal Agent capabilities.

The new model retains the 560 billion total parameters, 25 billion activated parameters, and 1 million Token context of the previous generation Ling-3.1-flash, but extends from text reasoning to native visual understanding. The earlier Ling-3.0-flash-VL already supported image and video inputs, but at a scale of only 124 billion total parameters, 5.5 billion activated parameters, and a context of 256,000 Tokens.

In official demos, the new model can build web pages from images, operate Blender to create 3D scenes, and read paper PDFs containing formulas and charts to turn the algorithms within into runnable code.

Ant's internal testing shows that the new model surpasses Claude Opus 5 (max) in three evaluations: document understanding, complex visual reasoning, and image-to-webpage generation. Among them, Vision2Web scored 85.86, compared with Opus 5's 80.20. On the third-party DesignArena image-to-frontend leaderboard, Ling-3.2-flash ranks 10th with 1238 Elo.

The model is now available for web experience and API, free for the first two weeks.